Add a gitea skill; point AGENTS.md at it

Forge usage was spread across AGENTS.md, docs/agents/issue-tracker.md, and whatever the agent remembered about `gh`. Move the how-to-run-`tea` half into a model-invoked skill so it loads on any issue/PR task, and leave AGENTS.md a single line of identity plus pointers. The skill carries only what `tea --help` cannot answer: `gh` is absent so the harness's `issue://`/`pr://` URIs error out, issues and PRs share one index space, rendered-box output needs `-o json`, `close` takes no `--comment`, labels never auto-create, bodies need a heredoc, and a read without `--comments` silently drops every comment (issue #123: 40 lines bare, 132 with the flag). Tracker conventions and triage label strings stay in docs/agents/, cited rather than copied.
docs: make every AGENTS.md cite code, not docs or issues (#113 )
2026-08-17 17:50:09 +07:00 · 2026-08-17 13:43:49 +07:00 · 2026-08-17 12:06:14 +07:00 · 2026-08-17 11:43:15 +07:00
8 changed files with 885 additions and 402 deletions
@@ -0,0 +1,75 @@
+---
+name: gitea
+description: Use for every forge operation in this repo — read, create, comment on, label, close, or search an issue; create, review, merge, or check out a PR; and whenever `gh`, `issue://`, or `pr://` fails or a ticket number is ambiguous. This repo's forge is self-hosted Gitea driven by `tea`, not GitHub.
+---
+
+# Gitea, not GitHub
+
+`origin` is the self-hosted Gitea instance `gitea.violetcrown.my.id`, repo
+`sulthan/mangaBookmark`. Everything past plain git goes through
+[`tea`](https://gitea.com/gitea/tea) (0.14.2 on this machine).
+
+**`gh` is not installed**, so the harness's `issue://<n>` and `pr://<n>` URIs
+error out (`GitHub CLI (gh) is not installed`, measured 2026-08-17). There is no
+fallback to add — read tickets with `tea`.
+
+`tea` infers the repo from `origin`; auth lives in `tea login`, never a
+`GH_TOKEN`. Your Gitea username comes from `tea login list` — `tea` has no `@me`.
+
+Flags are the environment's job: run `tea <command> --help` rather than trusting
+a remembered flag. This file carries only what `--help` will not tell you.
+
+## Commands
+
+| Job | Command |
+|---|---|
+| Read | `tea issue <n> --comments` / `tea pr <n> --comments` — `--comments` is not optional |
+| List | `tea issue list --state open\|closed\|all -o json --fields index,title,body,labels,state,author` |
+| Search | `tea issue list -k "<keyword>" -L "<label>" -A "<author>"` (`-K all` also searches PRs) |
+| Create | `tea issue create -t "..." -d "..."` (`-L`, `-a` optional) |
+| Comment | `tea comment <n> "..."` |
+| Label | `tea issue edit <n> --add-labels "..."` / `--remove-labels "..."` |
+| Close | `tea issue close <n>` / `tea pr close <n>` |
+| PR | `tea pr create --head <branch> --base main -t "..." -d "..."`, `tea pr checkout <n>`, `tea pr review <n>`, `tea pr merge <n>` |
+
+## Traps
+
+- **A bare read is a truncated read.** Without `--comments`, `tea issue <n>` and
+  `tea pr <n>` print the opening body and drop every comment silently — no
+  prompt, no marker, no hint that more exists (measured 2026-08-17: issue #123
+  prints 40 lines bare, 132 with `--comments`). The comments are where the
+  decisions live and the body is usually the stalest part of the ticket, so
+  **every read that exists to understand an issue or PR passes `--comments`**,
+  and understanding means body plus all comments plus whatever ticket they point
+  at. Comment count is `tea issue list --fields index,comments`, so a read that
+  shows fewer than that is incomplete. A PR's review comments are a second
+  stream: `tea pr review-comments <n>`.
+- **One index space for issues and PRs.** A bare `#42` may be either: try
+  `tea pr 42`, fall back to `tea issue 42`. Say which one you found.
+- **Output is rendered boxes**, not plain text. Anything you parse needs
+  `-o json`, plus `--fields` to keep the payload small. `tea pr create` prints
+  the PR URL on its last line.
+- **`close` takes no `--comment`.** Comment with `tea comment <n>`, then close.
+- **Gitea will not auto-create a label.** `tea labels list` first; missing one
+  gets `tea labels create --name "..." --color "#rrggbb"` before the `edit`.
+- **Multi-line bodies go through a heredoc**, never inline escapes:
+  ```bash
+  tea issue create -t "Title" -d "$(cat <<'EOF'
+  body line one
+
+  - acceptance criterion
+  EOF
+  )"
+  ```
+- **No sub-issue and no dependency command.** Gitea's API has issue
+  dependencies, `tea` does not expose them, so parentage and blocking live as
+  body lines — the shapes are in `docs/agents/issue-tracker.md`.
+
+## Conventions this repo layers on top
+
+Ticket bodies, wayfinding issues, and the PR-as-request-surface flag:
+`docs/agents/issue-tracker.md`. Triage label strings: `docs/agents/triage-labels.md`.
+A label named there still has to exist in the tracker before `--add-labels`.
+
+Finish a forge action by stating the number you touched and its state after —
+"commented and closed #71" — so the write is checkable without a second query.
@@ -11,7 +11,7 @@ to the tracker. You do not write the implementation — every line of ticket cod
 is written by a `ticket-implementer` subagent in its own git worktree. Reach for
 the editor yourself only for a merge conflict resolution.

-Ticket source and `tea` usage: `docs/agents/issue-tracker.md`.
+Ticket source and tracker conventions: `docs/agents/issue-tracker.md`. `tea` usage: skill `gitea`.

 ## 1. Collect the tickets

@@ -1,6 +1,6 @@
 # AGENTS.md

-Guidance for OpenCode (and Claude Code) working in this repo.
+Repo-wide guidance for coding agents.

 ## What this is

@@ -13,15 +13,18 @@ One backend, one `bookmarks` table: a `kind` column (`manga`|`novel`) splits the

 ## Hard constraints (drive design — don't violate)

+Nothing below is derivable from reading the code — it is why the code looks the
+way it does, plus dated measurements against services we don't control.
+
 Userscript targets **Violentmonkey**, so `GM_*` APIs available, but stay GM-free where plain web APIs suffice — keeps portability across engines:
 - **Avoid `GM_*` unless needed.** Prefer page `localStorage` over `GM_setValue`/`GM_getValue`, on-page UI over `GM_registerMenuCommand`, plain `fetch()` over `GM_xmlhttpRequest` for cross-origin.
 - Cross-origin `fetch()` work **only** against CORS-enabled backend. Manga sites `https://`, so backend **must be HTTPS** (else mixed-content block).
 - Every site is its **own origin with its own `localStorage`** — a shared remote store is the only way to unify bookmarks. Cloud sync required, not optional.
 - Userscript run in **isolated world**, so embedded API token safe from site's JS.
- Cloudflare's block on manga sites is **per-zone configuration plus request fingerprint, not IP reputation — and not reliably reproducible.** Verified 2026-07-26: plain `curl` from both CGNAT dev machine *and* deployed VPS got clean 200s with real HTML on both asurascans.com and demonicscans.org (homepage, series, chapter pages) — no interactive Turnstile challenge from either IP at test time. Contradicts earlier untested assumption CGNAT dev IP blocked; wasn't, at least this date. Treat "does curl work right now" as live, time-varying fact to re-check, not fixed property of machine — a Site can turn its protection on overnight, which is exactly what comix.to did on 2026-08-12. An earlier version of this line blamed "Cloudflare's bot scoring"; that was wrong. The 1-99 bot score is Enterprise Bot Management only and does not exist for a free-plan zone, and no per-IP request rate is documented as an input to challenge issuance — `docs/research/cloudflare-bot-scoring-and-poll-cadence.md`. Backend fetcher still needs graceful-degrade path for when challenged, and adapters should be **verified against live pages** (Playwright MCP, on-device devtools, direct probe) before finalizing, not assumed from single earlier test.
- **kagane.to, comix.to and novelfull.com are the exception to the above** — all three sit behind a Cloudflare JavaScript challenge no TLS fingerprint clears, so the backend polls them over CDP (`BROWSER_WS_URL`). When that's unset, kagane and comix are skipped entirely (a plain fetch would only retrieve a challenge page) while novelfull pages are still attempted over plain TLS — its challenge is a live time-varying fact and its cover bytes never need the browser. comix turned hostile on 2026-08-12 (#98): its cover host `static.comix.to` is gated too, so its cover bytes go through the browser as well, and its page is read as an in-tab `fetch()` of the series URL rather than a rendered DOM — comix is an SPA, and rendering costs ~65 requests for the same server-rendered HTML one fetch returns. The three other sites poll fine over plain TLS.
+- Cloudflare's block on manga sites is **per-zone configuration plus request fingerprint, not IP reputation — and not reliably reproducible.** Verified 2026-07-26: plain `curl` from both CGNAT dev machine *and* deployed VPS got clean 200s with real HTML on both asurascans.com and demonicscans.org (homepage, series, chapter pages) — no interactive Turnstile challenge from either IP at test time. Contradicts earlier untested assumption CGNAT dev IP blocked; wasn't, at least this date. Treat "does curl work right now" as live, time-varying fact to re-check, not fixed property of machine — a Site can turn its protection on overnight, which is exactly what comix.to did on 2026-08-12. An earlier version of this line blamed "Cloudflare's bot scoring"; that was wrong. The 1-99 bot score is Enterprise Bot Management only and does not exist for a free-plan zone, and no per-IP request rate is documented as an input to challenge issuance. Backend fetcher still needs graceful-degrade path for when challenged, and adapters should be **verified against live pages** (Playwright MCP, on-device devtools, direct probe) before finalizing, not assumed from single earlier test.
+- **kagane.to, comix.to and novelfull.com are the exception to the above** — all three sit behind a Cloudflare JavaScript challenge no TLS fingerprint clears, so the backend polls them over CDP (`BROWSER_WS_URL`). When that's unset, kagane and comix are skipped entirely (a plain fetch would only retrieve a challenge page) while novelfull pages are still attempted over plain TLS — its challenge is a live time-varying fact and its cover bytes never need the browser. comix turned hostile on 2026-08-12: its cover host `static.comix.to` is gated too, so its cover bytes go through the browser as well, and its page is read as an in-tab `fetch()` of the series URL rather than a rendered DOM — comix is an SPA, and rendering costs ~65 requests for the same server-rendered HTML one fetch returns. The three other sites poll fine over plain TLS.
 - **The CDP browser must look like a real browser, and stock headless images don't.** Measured 2026-08-08 against kagane.to, all from the same IP: `chromedp/headless-shell:stable` never cleared the challenge in 90s (`navigator.webdriver` true, empty plugin list, Chromium-branded client hints — suppressing `webdriver` alone changed nothing); `zenika/alpine-chrome` ships Chrome 124, refused outright; real Chrome with the default `--headless=new` UA never cleared, because the UA says `HeadlessChrome`; real Chrome with a stock UA **and** a non-UTC clock zone cleared in ~4s. Hence `chrome/` — a Debian image with `google-chrome-stable`, a version-derived UA, and `TZ`/`BROWSER_TZ`. Chrome reads the zone *name* through ICU from `/etc/localtime`'s symlink target, ignoring the file's contents, so mounting the host's `/etc/localtime` does **not** work; `/etc/timezone` is mounted instead.
- **The browser is not in the API stack and must not be put back.** It's its own compose unit (`chrome/docker-compose.yml`) on a second machine, reached over the tailnet — it held 471 MiB on a 1974 MiB swapless VPS, and a residential egress avoids the cloud-hosting-IP signature Bot Fight Mode documentedly challenges (ADR-0006; not a better "score" — free-plan zones have no score). Consequences that constrain code: `BROWSER_WS_URL` must be a tailnet **IP** (a MagicDNS name 500s at `/json/version`, same trap as the old Docker service name); the CDP port binds to the tailnet address only, since CDP authenticates nothing and that host has a real LAN; and the browser is on-demand (ADR-0005), so an unreachable or asleep one must degrade exactly as an unset `BROWSER_WS_URL` — plain-TLS libraries unaffected, kagane/comix logged and skipped, stored covers still served. Never add `chromedp.NoModifyURL`: discovery per fetch is what makes a restarted Chrome invisible.
+- **The browser is not in the API stack and must not be put back.** It's its own compose unit (`chrome/docker-compose.yml`) on a second machine, reached over the tailnet — it held 471 MiB on a 1974 MiB swapless VPS, and a residential egress avoids the cloud-hosting-IP signature Bot Fight Mode documentedly challenges (not a better "score" — free-plan zones have no score). Consequences that constrain code: `BROWSER_WS_URL` must be a tailnet **IP** (a MagicDNS name 500s at `/json/version`, same trap as the old Docker service name); the CDP port binds to the tailnet address only, since CDP authenticates nothing and that host has a real LAN; and the browser is on-demand, so an unreachable or asleep one must degrade exactly as an unset `BROWSER_WS_URL` — plain-TLS libraries unaffected, kagane/comix logged and skipped, stored covers still served. Never add `chromedp.NoModifyURL`: discovery per fetch is what makes a restarted Chrome invisible.
 - **UTC is the tell, not a country mismatch.** A UTC clock is the datacenter default, and the challenge refuses it; any real zone clears. Measured 2026-08-08, identical container, one Indonesian egress IP: UTC never cleared in 60s (twice), while `Asia/Jakarta` **and** `America/New_York` both cleared in 4s. An earlier note here claimed the zone had to match the egress IP's country — that was wrong, inferred from the host clock (`Asia/Bangkok`) rather than the measured egress. A second earlier claim, that Cloudflare "scores" a UTC clock, was also wrong: the measurement is real but the mechanism is not documented anywhere — Cloudflare publishes no timezone signal, and free-plan zones carry no score at all. `BROWSER_TZ` therefore needs a plausible zone, not a geolocated one.
 - **A challenged page needs the tab kept open.** The interstitial takes seconds to solve and only then writes clearance into the browser's shared cookie jar. Navigate-read-close never clears anything; `BrowserFetcher.run` holds one tab and re-reads until the payload arrives.

@@ -36,7 +39,7 @@ Two Violentmonkey userscripts (isolated world, per-site adapters, localStorage c
                                            on-demand Chrome, separate machine (chrome/)
 ```

-Two deployable units on two machines: the API stack (`docker-compose.yml` + `docker-compose.prod.yml`, on the VPS) and the browser (`chrome/docker-compose.yml`, on the home machine). They share nothing but `BROWSER_WS_URL` and update independently. Backend-specific architecture (packages, endpoints, poller, config env vars) lives in `backend/AGENTS.md`. Userscript-specific structure (adapters, retry queue, UI, live URL shapes) lives in `userscript/AGENTS.md`. Deploy order `DEPLOY.md` (§7 for the browser), redeploy `REDEPLOY.md` (§8 for the browser).
+Two deployable units on two machines: the API stack (`docker-compose.yml` + `docker-compose.prod.yml`, on the VPS) and the browser (`chrome/docker-compose.yml`, on the home machine). They share nothing but `BROWSER_WS_URL` and update independently. Backend-specific detail lives in `backend/AGENTS.md`, userscript-specific detail in `userscript/AGENTS.md`.

 ## Commands

@@ -64,28 +67,22 @@ Smoke test: `curl` endpoints with `Authorization: Bearer <token>`; confirm `OPTI

 ## Forge: Gitea, not GitHub

-`origin` is self-hosted Gitea instance (`gitea.violetcrown.my.id`), so **`gh` don't work here — use `tea` (Gitea CLI) for anything past plain git.** Common ones:
-
- Open PR: `tea pr create --head <branch> --base main --title "..." --description "..."`
- List / view / check out: `tea pr list`, `tea pr <n>`, `tea pr checkout <n>`
- Issues: `tea issue create`, `tea issue list`
- Auth lives in `tea login`, not `GH_TOKEN` env var.
-
-`tea` print output as rendered boxes rather than plain text; PR URL lands on last line.
+`origin` is self-hosted Gitea (`gitea.violetcrown.my.id`, repo `sulthan/mangaBookmark`), so **`gh` don't work here and the `issue://`/`pr://` URIs error out — drive the forge with `tea`.** How to run it — commands, traps, JSON output: skill `gitea`. Tracker conventions (ticket bodies, wayfinding, PR-as-request-surface flag): `docs/agents/issue-tracker.md`. Triage label strings: `docs/agents/triage-labels.md`.

 ## Design system

-Web UI + userscript panel follow **Cinder**, rules in `docs/design-system.md`
-— source of truth Claude Design project `BookmarkManager Web UI`
-(`969ac210-fe02-4c01-ae1b-9a271dcc779a`). Read it before touching
-`backend/internal/web/static/style.css`, `backend/internal/web/templates/*`, or userscript
-`TEMPLATE`/`CSS`. Core law: **ember means new chapter only** — no other
-state (busy, error, destruction) may use `--ember`; destruction gets
-`--danger`. No cards/corners/shadows, one `--measure: 760px` column, tokens
-only (never hardcode hex outside `:root`), both colour branches touched
-together. Any move that pulls series out of list (archive/finish/remove)
-must be confirm-gated via its own `.confirm-row`; only restore fires
-instantly.
+Web UI + userscript panel follow **Cinder**. Tokens are the `:root` block in
+`backend/internal/web/static/style.css`; that file, `backend/internal/web/templates/*`,
+and the userscript `TEMPLATE`/`CSS` are the only places it is expressed.
+Source of truth for the visual language is the Claude Design project
+`BookmarkManager Web UI` (`969ac210-fe02-4c01-ae1b-9a271dcc779a`).
+
+Core law: **ember means new chapter only** — no other state (busy, error,
+destruction) may use `--ember`; destruction gets `--danger`. No
+cards/corners/shadows, one `--measure: 760px` column, tokens only (never
+hardcode hex outside `:root`), both colour branches touched together. Any move
+that pulls a series out of the list (archive/finish/remove) must be
+confirm-gated via its own `.confirm-row`; only restore fires instantly.

 ## Security invariants

@@ -127,13 +124,19 @@ Review gate: auth, CORS, session, crypto, and the fetch gate are security-critic
 ## Comments

 Comment only if code alone can't carry info. Cost per read — must earn spot.
+Wrong comment worse than none: it misleads readers and measurably degrades
+LLM performance on the file. Missing comment costs little. Bias to fewer.

-Write for:
- Why not what. Tradeoffs, non-obvious decisions.
- Load-bearing detail looking incidental — say so if "simplify" breaks it.
+Docstring on public/exported surface — exception, near-always worth it.
+Contract only: what it takes, returns, throws, mutates; units; pre/post
+conditions. Not a restatement of the body. Skip on private/obvious.
+
+Inline — write for:
+- Why not what. Tradeoffs, non-obvious decisions, rejected alternatives.
+- heavy detail looking incidental — say so if "simplify" breaks it.
 - Non-local consequence, invisible from function alone.
- Wire format / encoding / interface contract — save callers re-deriving.
- Gotcha/workaround, with ref if exists.
+- Wire format / encoding / ordering / invariant — save callers re-deriving.
+- Gotcha/workaround, with ref (issue, RFC, vendor bug) if exists.
 - Domain/business rule not derivable from code.

 Skip:
@@ -142,24 +145,49 @@ Skip:
 - Banners, dividers, `// helpers`.
 - Change narration (`// fix bug`, `// as requested`, `// new impl`) — git's job.
 - Commented-out code — delete.
- TODO without concrete action.
+- TODO without concrete action + owner.
+- Narrating the plan you just reasoned through. Plan in prose or in your head;
+ship the code, not the transcript.
+- Anything restating a name that could be fixed by renaming instead.

-Style: one dense comment over function beats one per line inside. Tight, no worked example unless bug subtle. Wrong comment worse than none — update/delete on change. Default fewer — sparse+high-signal beats comprehensive.
+Staleness filter: if the comment describes something likely to change
+independently of this line, it will rot and start lying. Either anchor it to
+something stable, assert it in a test, or leave it out.

-Test: "competent reader get this from code in few sec?" Yes → skip. Needs detour through another file/spec/git-blame → write it.
+Style: one dense comment over a function beats one per line inside. Tight; no
+worked example unless the bug is subtle. On edit, update or delete stale
+comments in the code you touch — silence beats a lie.

-## Agent skills
+Test: "competent reader get this from code in a few sec?" Yes → skip.
+Needs detour through another file/spec/git-blame/external doc → write it.

-`AGENTS.md` is the single source of truth for agent guidance; every `CLAUDE.md` in this repo is a symlink to the `AGENTS.md` beside it. Edit `AGENTS.md`.
+## Writing an AGENTS.md

-### Issue tracker
+`AGENTS.md` is the single source of truth for agent guidance; every `CLAUDE.md`
+in this repo is a symlink to the `AGENTS.md` beside it. Edit `AGENTS.md`.

-Issues live as Gitea issues on `gitea.violetcrown.my.id` (`sulthan/mangaBookmark`), driven by the `tea` CLI — not `gh`. See `docs/agents/issue-tracker.md`.
+**Cite code, never docs, issues, or plans.** A spec, ADR, plan file, or Gitea
+issue records what was true when it was written and then goes stale silently;
+an agent that follows the pointer reads a decision that may already have been
+reversed. Code is the only source true at read time — cite a package, file,
+symbol, env var, or route. The sole non-code exception is a sibling
+`AGENTS.md`. If a doc holds a fact an agent needs, restate the fact here rather
+than linking to it.

-### Triage labels
+**State a fact in prose only if the code cannot answer it.** Split by
+derivability:

-Default five-role vocabulary, label strings unchanged (`needs-triage`, `needs-info`, `ready-for-agent`, `ready-for-human`, `wontfix`). See `docs/agents/triage-labels.md`.
+- *Structure* — packages, routes, env vars, columns, struct fields. Rots fast,
+  cheap to re-read. **Name the symbol, write nothing else.**
+- *Mechanism* — what a function does, how a flow proceeds. **Name the symbol
+  plus at most one line of orientation.**
+- *Rationale* — why it is this way, what a "simplify" would break, what was
+  tried and rejected. Not in the code and cannot be re-derived. **Write it out.**
+- *Measurement* — an observation against something we don't control. **Write it
+  out with the date**; a dated fact is honest, an undated one pretends to be
+  permanent.

-### Domain docs
-
-Single-context: one root `CONTEXT.md` plus `docs/adr/`, both created lazily. See `docs/agents/domain.md`.
+Restating mechanism in prose is how these files rot: the code changes, the
+paragraph doesn't, and the next agent trusts the paragraph. A pointer degrades
+more honestly — and every symbol you name must actually exist, since a dead
+pointer is a bug, not a stale sentence.
@@ -1,274 +1,268 @@
-Guidance for OpenCode (and Claude Code) working under `backend/`. See root `AGENTS.md` for the project-wide architecture diagram, hard constraints, and design system.
+Scope: `backend/`.

- **Backend** (`backend/`): stdlib `net/http` (handful routes, no framework) + Postgres over `jackc/pgx/v5` (pure Go, `CGO_ENABLED=0` -> static binary -> distroless/scratch image). Reverse proxy terminates TLS; Go service listens plain `:8080`.
-  Single binary, split into packages under `backend/internal/`: `store`
-  (Bookmark type, Postgres persistence, migration runner), `latest` (background
-  poller, site parsers, TLS fetcher), `session` (cookie signing, login
-  rate limiter), `httpmw` (Auth/Gzip/CORS middleware), `api` (JSON
-  bookmark handlers), `userscript` (userscript-serving handler), `web`
-  (browser UI handler + `templates/` + `static/`, `go:embed`-ed).
-  `backend/main.go` is the composition root — the only place that wires
-  packages together into `newRouter`. Root-level `*_test.go` hold
-  integration tests that exercise the full router; unit tests for a
-  package live beside it under `internal/`.
- **Schema is migration-owned.** `internal/store/migrations/*.sql` is
-  `go:embed`-ed and applied on every start by `store.migrate`: one numbered
-  file per change, one transaction each, versions recorded in
-  `schema_migrations`. Files are **append-only** — editing an applied one
-  changes nothing on a database that already ran it. No column probing, no
-  data-fixup migrations: both were SQLite-era machinery and are gone.
- **Tests need Docker.** `internal/pgtest` starts one `postgres:17-alpine`
-  container per test binary (`TestMain` -> `pgtest.Main`) and hands each test
-  its own database (`pgtest.URL(t)`). A package whose tests touch the store
-  must have that `TestMain`.
- **Reader-owned store, four tables.** `readers` is keyed by Discord user ID
-  and carries the SHA-256 of the Reader's userscript credential plus a
-  `token_epoch` (issue #24). Credentials are derived, never stored: `token.Token(TOKEN_KEY, discord_id, epoch)` (HMAC, `internal/token`), and only its SHA-256 sits in `readers.token_sha256`, so install URLs can be rebuilt after any restart while a database leak yields nothing but hashes. The seed creates the **owner** row at startup; its epoch-0 hash is refreshed on every start **only while the row has never been rotated**, so a restart can never resurrect a rotated-away credential. Every other row is created by that Reader's own first login (`Store.EnsureReader`, idempotent on `discord_id`, and it never rewrites an existing row's hash). Rotation is `Store.RotateToken` (epoch bump + hash rewrite in one transaction), driven by the web UI.
-  `series` keyed `(site, series_id)`
-  (`asura`|`demonic`|`comix`|`kagane`|`novelfull`|`lightnovelworld`) owns the
-  shared facts — title, cover, canonical URL, `kind` (`manga`|`novel`),
-  Latest Chapter, `latest_checked_at`, and the Sighting pair
-  `latest_sighted_at`/`latest_raised_by` (issue #103) — and `bookmarks` holds only what
-  differs between readers: progress, favourite, lifecycle bucket,
-  `updated_at`. A bookmark is keyed `(reader_id, site, series_id)` — no
-  surrogate id; the wire `key` is derived as `site:series_id` on read — and
-  every store read/write is scoped to the reader it names. Auth resolves the
-  acting Reader from the presented credential (`httpmw.Auth`) and nothing
-  else — there is no unauthenticated-by-Reader route and no global token; the
-  reader id travels in the request context. Sync **last-write-wins**; the wire format
-  stays flat (ADR-0004). `Store.Upsert` decomposes one flat body across two
-  tables and enforces the ownership rule: client `title`/`series_url`/`cover`
-  are written only when the series row is new (ADR-0003).
- **Endpoints:** `GET /bookmarks`, `PUT /bookmarks/{key}` (upsert; see `updated_at` rule below), `DELETE /bookmarks/{key}`, `GET /healthz` (no auth).
- **Web UI:** same binary serve the browser UI on a second
-  hostname — `GET /` (list, or login page when no session),
-  `GET /auth/discord` + `GET /auth/discord/callback` (Discord OAuth,
-  ADR-0002), `POST /logout`, `GET /static/*`, htmx fragment endpoints
-  under `/ui/*`. Templates + assets `go:embed`-ed under
-  `backend/internal/web/`, so `backend/Dockerfile` must copy the whole
-  `internal/` tree, not just `*.go`. Sessions are rows in the `sessions`
-  table: the cookie carries only an opaque id, looked up (and expiry-
-  checked) on every request, and deleting the row revokes the session.
-  Guild membership *is* registration (issue #27): `discordCallback` gates on
-  membership (and `DISCORD_REQUIRED_ROLE` when set) and then calls
-  `Store.EnsureReader`, so a refusal creates nothing and a returning Reader
-  reuses their row. The owner is the only Reader with administrative reach:
-  `POST /readers/{id}/revoke` (404 for anyone else) drops that Reader's
-  sessions, and the `readers` panel renders only on the owner's page.
-  A Reader with no bookmarks at all sees `listView.Fresh`, whose empty state
-  offers both install links instead of describing a filter.
-  UI mutations read-modify-write
-  through `Store.Get` + `Store.Upsert` so `updated_at` rule stays one
-  place. See `docs/superpowers/specs/2026-07-25-web-ui-design.md`.
-  **Design-tool caveat:** templates link `/static/style.css` root-absolutely
-  (correct — served from `/`), but impeccable detector resolves
-  stylesheet href with `path.resolve(fileDir, href)`, drops directory
-  on leading `/` and silently skip file. Relative href don't help
-  either: template's directory isn't its served path. So
-  `detect.mjs backend/internal/web/templates` reports **false clean** —
-  always pass `backend/internal/web/static` too. One finding there,
-  `overused-font` on "Instrument Serif", deliberate identity choice, not debt.
- **Every action that moves series out of list is confirm-gated.**
-  Archive, finish, remove each open own `.confirm-row` disclosure
-  (`toggleConfirmRow(key, kind)` in `filter.js`, `kind` ∈
-  `archive|finish|remove`); restore fire instantly since it's the reversal.
-  Remove's row wear ember wash, two reversible ones wear `.calm` grey.
-  `--ember` stay reserved for new-chapter signal: busy bar and inline
-  error use `--mute`.
- **Latest-chapter poller:** one goroutine per Site (a Poll Lane, issue #100),
-  each re-checking that Site's bookmarked series' newest published chapter from
-  backend's own network access, so `latest_chapter` stays fresh when the user
-  isn't browsing. Second, parallel signal — the userscript keeps its own
-  `maybeCaptureLatestOnSeriesPage`/`backgroundRefreshLatest` schedule, and its
-  `reportLatestChapter` PUTs every read, unchanged numbers included, because an
-  unchanged read is exactly the Sighting worth deferring a Poll on (#103).
-  Two independent clocks: per-series rest (`series.latest_checked_at`,
-  enforced by `Store.DueForLatestCheck`'s WHERE clause — `now - Rest`) and
-  per-Lane gap (the Lane sleeping between fetches, `effectiveGap`). Both live
-  in the Site registry (`internal/latest/sites.go`), not config: the five env
-  knobs that used to size a shared pace are gone.
-  The poller walks **Series, not Bookmarks** — a series referenced by several
-  bookmarks is fetched once per cycle, and the due queue orders
-  `reader_count DESC, latest_checked_at ASC` (ADR-0003). Series row stamped
-  *before* fetch so broken series wait out the rest instead of retrying
-  every tick; found chapter written straight to the series row via
-  `Store.SetLatestChapter`, so a bookmark's `updated_at` — and the list
-  order — is never touched.
-  **Sightings** (issue #103, ADR-0011) let a Reader's own page read defer a
-  Poll: `Store.RecordSighting` — called by the PUT handler *before* the Upsert,
-  because the raise test needs the row as it stands — stamps
-  `series.latest_sighted_at` and, when the report raises the stored number,
-  names its Reader in `series.latest_raised_by`. The due query's HAVING clause
-  is where deferral lives: a Series is skipped only while it has exactly one
-  Bookmark, was sighted within one Rest, and is under the ceiling
-  (`sightingCeilingRests`, six of that Site's rests) since its last Poll. So a
-  shared Series is never deferred, and no Series goes six hours unpolled
-  whatever arrives. `checkOne` judges the named Reader off the comparison it
-  already makes: a lower number is a contradiction (logged with the Reader and
-  both numbers), the same number an agreement, a higher number the Site
-  publishing and neither — that last one clears the attribution instead, since
-  the value the Poll then stores is its own and a later retraction is not the
-  Reader's fault. Three contradictions
-  (`store.SightingDisagreementLimit`) stop that Reader deferring — their
-  reports still write the Latest Chapter — and twenty consecutive agreements
-  (`store.SightingAgreementsToClear`) forgive them, as does the owner's
-  clear-marks control. Deferral is recomputed from live facts every round, so
-  nothing needs invalidating when a Series gains a second Bookmark; the one
-  input read earlier is the Reader's marks, checked when the Sighting is
-  recorded, so crossing the threshold or being cleared takes effect from that
-  Reader's next Sighting and the standing already bought lasts out its rest.
-  Refusals and browser loss are Lane-local: two `errChallengeHeld` in one pass
-  stop that Site for `refuseBackoff` (15m) while other Lanes continue; an
-  `errBrowserInterrupted` (remote Chrome restart) sets a shared Poller flag
-  that makes the other browser Lanes skip their passes for the same 15m, so a
-  restarting Chrome doesn't stamp one Series per Lane per pass — after the
-  window the flag decays and they probe again. Browser Lanes wake Chrome only
-  when 5+ Series are due or one has waited 15m (ADR-0005 on-demand browser),
-  and cover work (both healing a stored source URL and filling a blank from
-  the series page) runs in the background so a slow CDN can't consume a
-  Lane's gap.
-  A refusal is only ever the challenge *page*: `isInterstitial` matches the
-  orchestration path `/cdn-cgi/challenge-platform/h/`, never the bare prefix.
-  Cloudflare injects `/cdn-cgi/challenge-platform/scripts/jsd/main.js` into
-  ordinary 200 pages once a zone turns JS detections on, which demonic did on
-  2026-08-16 — the prefix match then read every real demonic page as a refusal
-  and parked that Lane in 15m backoff while plain TLS was returning the full
-  series page.
-  Fetches use `bogdanfinn/tls-client` with Chrome profile as defence in depth
-  against fingerprint-based blocking; any failure log and skip. kagane, comix
-  and novelfull sit behind Cloudflare JavaScript challenges the TLS client
-  can't clear, so they are fetched over CDP via `BROWSER_WS_URL`; kagane and
-  comix are simply not polled when that's unset, while novelfull falls back to
-  a plain-TLS attempt — its challenge is a live time-varying fact, and its
-  cover bytes never need the browser. comix's browser read is an in-tab
-  `fetch()` of the Series URL, not a DOM render: it is an SPA, so rendering
-  costs ~65 requests for the same server-rendered HTML one fetch returns
-  (measured 2026-08-12, issue #98). See
-  `docs/superpowers/specs/2026-07-26-server-latest-chapter-polling-design.md`.
-  The poller's series write is a single-column UPDATE
-  (`Store.SetLatestChapter`), not a read-modify-write of the whole bookmark:
-  it cannot revert read progress or move `updated_at`, so the old
-  stale-re-read race is gone with the Get+Upsert flow.
- **Covers are acquired at creation, then served from our own origin
-  (ADR-0007):** the first Bookmark of a Series fires `Store.OnSeriesCreated`,
-  which `latest.Acquirer` turns into one series-page fetch yielding both the
-  Latest Chapter and the cover URL; the bytes then go through
-  `latest.CoverBytesFetcher` into `Store.SetSeriesCover`. It runs in a
-  goroutine — the Reader's PUT must neither block on a Site nor fail with one
-  — and every failure is logged and dropped, leaving the Bookmark intact. The
-  wire's `cover` is the absolute `PUBLIC_BASE_URL + /covers/{sha256}` once
-  bytes exist and `""` before, never an address that 404s. `GET /covers/{addr}`
-  is public and uncredentialed: the userscript renders it on a Site's origin,
-  where no cookie or token of ours travels. A client-sent `cover` is decoded
-  and discarded, permanently (ADR-0004 compatibility).
-  Browser-backed Sites join the same pipeline (issue #62, extended to comix by
-  #98): kagane and comix pages *and* cover bytes go through the browser sidecar
-  (nothing falls back to a plain fetch, which would only retrieve a challenge
-  page), while novelfull needs the browser only for its HTML — the cover URL
-  comes out of the browser-fetched page and the bytes go over plain TLS. With
-  no browser configured, kagane and comix Covers are simply absent; novelfull
-  still gets one — at creation and on the poll — when its page body happens to
-  answer a plain request (the challenge is a live time-varying fact). comix
-  cover bytes must arrive by direct navigation, not an in-page fetch: its
-  Series page sets `cross-origin-embedder-policy: require-corp`, which fails a
-  page-context fetch of `static.comix.to`. The old kagane-only
-  serving path (`/img/kagane/{id}`, template rewrite, `CoverFetcher`) is gone
-  (issue #63): the one public route serves every Site.
- **`updated_at` drives list order, so moves only on real reading progress:** server apply its timestamp when row new or `last_chapter_num` changes, else keep stored value — favouriting series or recording newly published chapter must not reorder list. `PUT` therefore returns row **as stored**, clients must adopt that response rather than own payload. See `plans/2026-07-25-bookmark-list-favorites-design.md` §4.
- **Lifecycle buckets:** `status` on each bookmark is `reading` | `archived` |
-  `finished`, orthogonal to `favorite`. Archived and finished appear only in
-  own tab — not in All, Updated, Favourites, or recent strip. Poller keeps
-  checking archived series and skip finished ones. `finished` settable
-  only from web UI; `PUT /bookmarks/{key}` reject it with 400.
-  **Empty incoming status means "keep stored one"** — resolved on the
-  `VALUES` side of `Store.Upsert`, not conflict clause, since
-  `excluded.*` is post-evaluation row and default applied there would
-  wipe bucket on every PUT from client that predates column. See
-  `docs/superpowers/specs/2026-07-27-status-buckets-design.md`.
- **Config via env:** `TOKEN_KEY` (derives every Reader's userscript credential;
-  required), `OWNER_DISCORD_ID` (seeds the owner Reader — the administrator and
-  the owner of every pre-registration bookmark; required),
-  `ALLOWED_ORIGINS` (comma list),
-  `DATABASE_URL` (Postgres connection URL, required — no default),
-  `COVER_DIR` (required filesystem volume for content-addressed Cover bytes),
-  `PUBLIC_BASE_URL` (required origin this deployment answers on, trailing
-  slash trimmed; every Cover URL on the wire is built from it, absolute
-  because the userscript renders on a Site's origin — ADR-0007),
-  `PORT` (default `8080`), `DISCORD_CLIENT_ID`/`_CLIENT_SECRET`/`_GUILD_ID`/
-  `_REDIRECT_URI` (required; Discord OAuth for the browser UI),
-  `DISCORD_REQUIRED_ROLE` (optional role gate, empty by default),
-  `DISCORD_API_BASE` (default `https://discord.com/api/v10`),
-  `LATEST_CHAPTER_POLL_ENABLED` (background latest-chapter poller kill
-  switch, default on). Pace is per Site in the registry (issue #100): every
-  Site rests an hour and gaps ten seconds, a Site with more eligible Series
-  than 360 tightens its own gap toward the 1s floor, and browser Lanes wake
-  Chrome only on demand (ADR-0005). The `_COOLDOWN`/`_BROWSER_COOLDOWN`/
-  `_INTERVAL`/`_BATCH`/`_STAGGER` knobs that used to size a shared pace are
-  gone. The 1h rest for browser Sites is safe on documented grounds: a
-  challenged page costs seconds of a serialized single-tab browser, free-plan
-  zones have no bot score and no published per-IP rate input, and
-  `cf_clearance` expires in 30 minutes so every cadence at or above 1h
-  re-solves anyway —
-  `docs/research/cloudflare-bot-scoring-and-poll-cadence.md`.
-  `USERSCRIPT_PATH` and `NOVEL_USERSCRIPT_PATH` (files served at
-  `/u/{token}/manga-bookmark.user.js` and `/u/{token}/novel-bookmark.user.js`,
-  defaults `/userscript/manga-bookmark.user.js` and
-  `/userscript/novel-bookmark.user.js`, both supplied by bindmount; the
-  `__API_TOKEN__` placeholder inside them is substituted with the requesting
-  Reader's credential at serve time).
-  `BROWSER_WS_URL` (CDP endpoint of the browser, which runs on a **separate
-  machine** and is reached over the tailnet — ADR-0006, `chrome/docker-compose.yml`.
-  Used by the poller for kagane, comix and novelfull page fetches and by the
-  cover pipeline for kagane's and comix's image bytes (the browser is the only
-  route that clears the challenge those two serve their covers behind); unset —
-  the default — disables browser polling and leaves kagane and comix Covers
-  blank until stored bytes
-  exist. Must be a tailnet IP, never a hostname: Chrome's DevTools handler 500s
-  `/json/version` for any Host that isn't an IP or `localhost`).
- **No per-Site cover path (issue #63):** every Cover — all six Sites — is
-  served by the one public `GET /covers/{addr}` route from content-addressed
-  bytes. There is no proxy, no per-Site rewrite, no second place that decides
-  a Cover's renderable address: the wire `cover` is it. The only place a Site
-  name still appears in cover code is the extraction module (`latest`), where
-  kagane's and comix's image URLs are claimed by `browserOnlyCoverURL` — kagane
-  answers a plain fetch with a challenge and
+Each entry names the code that holds the truth — read that for *what it does*.
+The prose here is only what code cannot tell you: rationale, rejected
+alternatives, dated measurements, and invariants a plausible refactor would
+silently break.
+
+### Layout
+
+`backend/main.go` → `newRouter` is the composition root, the only place
+packages are wired. Packages under `backend/internal/`: `store`, `latest`,
+`session`, `httpmw`, `api`, `userscript`, `web`, `token`, `pgtest`. Root-level
+`*_test.go` exercise the full router; unit tests live beside their package.
+
+Not visible from any single file: stdlib `net/http` with no framework,
+Postgres over `jackc/pgx/v5`, `CGO_ENABLED=0` static binary into a distroless
+image, TLS terminated by the reverse proxy so the service listens plain `:8080`.
+
+### Schema — `internal/store/migrations/*.sql`, run by `store.migrate`
+
+- Migration files are **append-only**. Editing an applied one changes nothing
+  on a database that already recorded its version in `schema_migrations`, so
+  the fix silently applies to new deployments only.
+- No column probing, no data-fixup migrations. Both were SQLite-era machinery
+  and were removed deliberately — don't reintroduce either.
+
+### Tests need Docker — `internal/pgtest`
+
+`pgtest.Main` from `TestMain` starts one `postgres:17-alpine` per test binary;
+`pgtest.URL` hands each test its own database. A package whose tests touch the
+store must have that `TestMain` or it has no database at all.
+
+### Reader-owned store — `internal/store`, `internal/token`
+
+Four tables; shape is in the migrations, behaviour in `Store`'s methods.
+
+- **Credentials are derived, never stored.** `token.Token(TOKEN_KEY, discord_id, epoch)`
+  is an HMAC; only its SHA-256 reaches `readers.token_sha256`. So install URLs
+  can be rebuilt after any restart, and a database leak yields nothing usable.
+- **The owner's epoch-0 hash is refreshed at startup only while the row has
+  never been rotated.** Drop that condition and a restart resurrects a
+  rotated-away credential.
+- `Store.EnsureReader` never rewrites an existing row's hash — a returning
+  Reader's login must not invalidate their installed scripts.
+- **Every read and write is scoped to the acting Reader**, resolved from the
+  presented credential by `httpmw.Auth` and carried in the request context.
+  There is no unauthenticated-by-Reader route and no global token.
+- **`series` holds what readers share, `bookmarks` only what differs.** A
+  bookmark key is `(reader_id, site, series_id)` with no surrogate id; the wire
+  `key` is derived as `site:series_id` on read.
+- `Store.Upsert` splits one flat body across both tables and enforces the
+  ownership rule: client `title`/`series_url`/`cover` are written **only when
+  the series row is new**, so one reader cannot retitle a shared series.
+- Sync is last-write-wins and the wire format stays flat — clients depend on
+  both; neither is an implementation detail to tidy up.
+
+### Web UI — `internal/web`
+
+Routes, templates and assets are all in that package; `AdminPatterns()` and
+`adminRoutes()` enumerate the privileged ones.
+
+- **`backend/Dockerfile` must copy the whole `internal/` tree**, not just
+  `*.go`: templates and static assets are `go:embed`-ed from
+  `internal/web/`.
+- **Guild membership *is* registration.** `discordCallback` gates on membership
+  (plus `DISCORD_REQUIRED_ROLE` when set) and only then calls
+  `Store.EnsureReader`, so a refusal creates nothing.
+- Sessions are rows, not signatures: the cookie carries an opaque id and
+  expiry is checked on lookup, which is what makes deleting the row an instant
+  revocation.
+- UI mutations go through `Store.Get` + `Store.Upsert` so the `updated_at` rule
+  below stays in exactly one place.
+- `listView.Fresh` exists because a Reader with no bookmarks at all needs
+  install links, not an empty-filter message.
+- **Design-tool caveat:** `detect.mjs backend/internal/web/templates` reports a
+  **false clean**. Templates link `/static/style.css` root-absolutely (correct —
+  it is served from `/`), but the detector resolves hrefs with
+  `path.resolve(fileDir, href)`, which drops the directory on a leading `/` and
+  skips the file silently; a relative href doesn't help either, since a
+  template's directory isn't its served path. Always pass
+  `backend/internal/web/static` too. The one finding there, `overused-font` on
+  "Instrument Serif", is a deliberate identity choice, not debt.
+
+### Confirm gating — `internal/web/static/filter.js`, `toggleConfirmRow(key, kind)`
+
+Every action that pulls a series out of the list (`archive|finish|remove`) opens
+its own `.confirm-row`; restore fires instantly because it is the reversal.
+Remove wears the ember wash, the two reversible ones wear `.calm` grey.
+**`--ember` is reserved for the new-chapter signal** — the busy bar and inline
+errors must use `--mute`, or the one colour that means "something to read"
+stops meaning it.
+
+### Latest-chapter poller — `internal/latest`, Site registry in `sites.go`
+
+One goroutine per Site (a Poll Lane) re-checks that Site's bookmarked series
+from the backend's own network position, so `latest_chapter` stays fresh while
+nobody is browsing. The userscript's `reportLatestChapter` is a second,
+parallel signal — it PUTs every read, unchanged numbers included, because an
+unchanged read is exactly the Sighting worth deferring a Poll on.
+
+- **Pace lives in the Site registry, not config.** Two clocks: per-series rest
+  (`series.latest_checked_at`, enforced in `Store.DueForLatestCheck`'s WHERE)
+  and per-Lane gap (`effectiveGap`). The five env knobs that used to size one
+  shared pace are gone; don't add them back.
+- **The poller walks Series, not Bookmarks** — a series several readers hold is
+  fetched once per cycle, and the due queue orders `reader_count DESC,
+  latest_checked_at ASC` so the widely-read ones win contention.
+- **The series row is stamped *before* the fetch**, so a permanently broken
+  series waits out its rest instead of being retried every tick.
+- `Store.SetLatestChapter` is a single-column UPDATE, deliberately not a
+  read-modify-write of the bookmark: it therefore cannot revert read progress
+  or move `updated_at`. The old stale-re-read race died with the Get+Upsert
+  flow — don't restore one here.
+
+**Sightings** (`Store.RecordSighting`, the due query's HAVING clause,
+`latest.checkOne`) let a Reader's own page read defer a Poll.
+
+- Recorded by the PUT handler **before** the Upsert, because the raise test
+  needs the row as it stands.
+- A Series is deferred only while it has exactly one Bookmark, was sighted
+  within one Rest, and is under `sightingCeilingRests` since its last Poll — so
+  a shared Series is never deferred and nothing goes six hours unpolled
+  whatever arrives.
+- A *higher* report clears the attribution rather than crediting it: the value
+  the Poll then stores is its own, so a later retraction isn't the Reader's
+  fault.
+- `store.SightingDisagreementLimit` contradictions stop a Reader deferring —
+  their reports still write the Latest Chapter — and
+  `store.SightingAgreementsToClear` agreements forgive them, as does the
+  owner's clear-marks control.
+- Deferral is recomputed from live facts each round, so nothing needs
+  invalidating when a Series gains a second Bookmark. The one input read
+  earlier is the Reader's marks, so crossing or clearing a threshold takes
+  effect from their next Sighting and the standing already bought lasts out its
+  rest.
+
+**Refusals and browser loss are Lane-local.** Two `errChallengeHeld` in a pass
+stop that Site for `refuseBackoff` while other Lanes continue. An
+`errBrowserInterrupted` (remote Chrome restarted) sets a shared Poller flag so
+the *other* browser Lanes skip their passes for the same window — otherwise a
+restarting Chrome stamps one Series per Lane per pass, burning rests on
+failures. The flag decays and they probe again.
+
+- **`isInterstitial` matches the orchestration path
+  `/cdn-cgi/challenge-platform/h/`, never the bare prefix.** Cloudflare injects
+  `/cdn-cgi/challenge-platform/scripts/jsd/main.js` into ordinary 200 pages
+  once a zone turns JS detections on, which demonic did on 2026-08-16: the
+  prefix match read every real demonic page as a refusal and parked the Lane in
+  backoff while plain TLS was returning full series pages.
+- Fetches use `bogdanfinn/tls-client` with a Chrome profile as defence in depth
+  against fingerprint blocking; any failure logs and skips.
+- kagane, comix and novelfull sit behind Cloudflare JS challenges the TLS
+  client can't clear, so they go over CDP (`BROWSER_WS_URL`). kagane and comix
+  are simply not polled when it's unset — a plain fetch would only retrieve a
+  challenge page — while novelfull still attempts plain TLS, because its
+  challenge is a live time-varying fact and its cover bytes never need a browser.
+- **comix's browser read is an in-tab `fetch()` of the Series URL, not a DOM
+  render.** It is an SPA: rendering cost ~65 requests for the same
+  server-rendered HTML one fetch returns (measured 2026-08-12).
+- Browser Lanes wake Chrome only when 5+ Series are due or one has waited 15m,
+  and cover work runs in the background so a slow CDN can't eat a Lane's gap.
+
+### Covers — `Store.OnSeriesCreated`, `latest.Acquirer`, `latest.CoverBytesFetcher`, `Store.SetSeriesCover`
+
+Acquired once when the first Bookmark of a Series is created, then served from
+our own origin by the public `GET /covers/{addr}`.
+
+- Acquisition runs in a goroutine: the Reader's PUT must neither block on a
+  Site nor fail with one. Every failure is logged and dropped, leaving the
+  Bookmark intact.
+- The wire `cover` is the absolute `PUBLIC_BASE_URL + /covers/{sha256}` once
+  bytes exist and `""` before — **never an address that 404s**. Absolute
+  because the userscript renders it on a Site's origin.
+- `GET /covers/{addr}` is public and uncredentialed by design: no cookie or
+  token of ours may travel to a Site's origin.
+- A client-sent `cover` is decoded and discarded, permanently — wire
+  compatibility, not an oversight.
+- **One route serves all six Sites.** No proxy, no per-Site rewrite, no second
+  place that decides a renderable address: the wire `cover` is it. Templates
+  render `.Cover` and nothing else. The old kagane-only serving path
+  (`/img/kagane/{id}` plus a template rewrite) is gone; don't reintroduce a
+  per-Site route because one Site's CDN misbehaves.
+- The only Site names left in cover code are in `browserOnlyCoverURL`
+  (`internal/latest`): kagane answers a plain fetch with a challenge *and*
  `cross-origin-resource-policy: same-origin`, and `static.comix.to` answers
-  one with the same Cloudflare challenge its pages serve;
-  every other Site's CDN answers plain TLS. Templates render `.Cover` — the
-  wire value — never anything else.
- **Web UI also owns:** session-gated `GET /install/{manga,novel}-bookmark.user.js`
-  (renders the bindmounted script with the acting Reader's derived credential
-  substituted in — the credential never appears in page markup, the address
-  bar, or a redirect; `?download=1` adds `Content-Disposition: attachment` for
-  mobile Violentmonkey, which ignores a `.user.js` navigation) and
-  `POST /rotate-token` (atomic epoch bump + hash
-  rewrite; invalidates every installed copy, so the panel warns to reinstall
-  on all devices).
- **Owner-only admin page (`internal/web/admin.go`, issue #102):** `GET /admin`
-  carries the Reader roster (sessions, Sighting counters, `POST
-  /readers/{id}/revoke` and `POST /readers/{id}/clear-marks`) and Poll Lane
-  status (`GET /ui/admin/lanes`, self-refreshing every 30s). Every route that
-  reaches past the acting Reader is listed in `adminRoutes()` and wrapped in
-  `requireOwner` at registration — add a route there, not a check inside a
-  handler; `web.AdminPatterns()` is what the gate test walks. A non-owner gets
-  404, never 403. Lane figures come from the running poller through the
-  `web.LaneReporter` seam (`latest.Poller.LaneStatus`), never from a table: a
-  nil reporter or a Lane that has not finished a pass renders "no data yet"
-  rather than zeroes. `main.newRouter` takes the reporter as an interface and
-  converts a nil `*Poller` to a nil interface — a typed nil would make the page
-  claim a poller exists.
-  The one owner comparison left outside `requireOwner` is in `index`
+  with the same challenge its pages serve. Every other Site's CDN answers plain
+  TLS.
+- **comix cover bytes must arrive by direct navigation, not an in-page fetch:**
+  its Series page sets `cross-origin-embedder-policy: require-corp`, which
+  fails a page-context fetch of `static.comix.to`.
+- With no browser configured, kagane and comix Covers are simply absent;
+  novelfull still gets one whenever its page answers a plain request.
+
+### `updated_at` drives list order — `Store.Upsert`
+
+The server applies its own timestamp only when the row is new or
+`last_chapter_num` changes, else it keeps the stored value. **Favouriting a
+series, or a newly published chapter arriving, must not reorder the list** —
+only real reading progress moves a row. Consequently `PUT` returns the row **as
+stored** and clients must adopt that response rather than their own payload.
+
+### Lifecycle buckets — `status` on each bookmark
+
+`reading` | `archived` | `finished`, orthogonal to `favorite`. Archived and
+finished appear only in their own tab, never in All, Updated, Favourites or the
+recent strip. The poller keeps checking archived series and skips finished ones.
+
+- `finished` is settable only from the web UI; `PUT /bookmarks/{key}` rejects
+  it with 400.
+- **An empty incoming status means "keep the stored one"**, and it is resolved
+  on the `VALUES` side of `Store.Upsert`, not in the conflict clause:
+  `excluded.*` is the post-evaluation row, so a default applied there would
+  wipe the bucket on every PUT from a client predating the column.
+
+### Config — `Config` / `loadConfig` / `loadLatestPoll` in `backend/main.go`
+
+That function is the complete list of env vars, their defaults, and which are
+required. What it can't tell you:
+
+- `PUBLIC_BASE_URL` must be an absolute origin because every Cover URL on the
+  wire is built from it and the userscript renders on a Site's origin.
+- `BROWSER_WS_URL` **must be a tailnet IP, never a hostname** — Chrome's
+  DevTools handler 500s `/json/version` for any Host that isn't an IP or
+  `localhost`. Unset (the default) disables browser polling.
+- `USERSCRIPT_PATH` / `NOVEL_USERSCRIPT_PATH` are bindmounted files; the
+  `__API_TOKEN__` placeholder inside them is substituted with the requesting
+  Reader's credential at serve time.
+- Pace is per Site in the registry, not env. The
+  `_COOLDOWN`/`_BROWSER_COOLDOWN`/`_INTERVAL`/`_BATCH`/`_STAGGER` knobs are
+  gone on purpose.
+- The 1h rest for browser Sites is safe on documented grounds: a challenged
+  page costs seconds of a serialized single-tab browser, free-plan zones carry
+  no bot score and no published per-IP rate input, and `cf_clearance` expires
+  in 30 minutes, so every cadence at or above 1h re-solves anyway.
+
+### Userscript install & rotation — `internal/userscript`, `internal/token`
+
+Session-gated `GET /install/{manga,novel}-bookmark.user.js` renders the
+bindmounted script with the acting Reader's derived credential substituted in,
+so the credential never appears in page markup, the address bar, or a redirect.
+`?download=1` adds `Content-Disposition: attachment` for mobile Violentmonkey,
+which ignores a `.user.js` navigation. `POST /rotate-token` is an atomic epoch
+bump plus hash rewrite and invalidates every installed copy — the panel must
+keep warning to reinstall on all devices.
+
+### Owner-only admin — `internal/web/admin.go`
+
+- **Every route reaching past the acting Reader is listed in `adminRoutes()`
+  and wrapped in `requireOwner` at registration** — add it there, not as a
+  check inside a handler; `web.AdminPatterns()` is what the gate test walks. A
+  non-owner gets 404, never 403.
+- The one owner comparison left outside the gate is in `index`
  (`view.Owner = readerID == h.store.OwnerID()`): it gates a link, not an
  endpoint, so it is a rendering decision a registration-time wrapper cannot
-  express — do not "unify" it into the gate.
-  A Lane pass that returns before computing its figures (refusal backoff,
-  sidecar down) carries the previous pass's due count and gap forward rather
-  than recording zeroes; a Lane that has never reached a pace renders no gap at
-  all. `Checked` next to `Due` is what separates a stopped Lane from a quiet
-  one, so neither figure may be dropped from the row.
-  Due-without-Checked is *not* by itself a stall: a browser Lane under both
-  wake thresholds sets `LaneState.Asleep` at the on-demand gate and renders
-  "browser asleep" instead of "not checking", and never counts toward
-  `Attention`. That is the commonest healthy state for kagane, comix and
-  novelfull — one due Series, nothing checked — so spending the stall mark on
-  it would train the owner to ignore the mark that matters.
+  express. Do not "unify" it into the gate.
+- Lane figures come through the `web.LaneReporter` seam
+  (`latest.Poller.LaneStatus`), never a table. `main.newRouter` takes the
+  reporter as an interface and converts a nil `*Poller` to a nil interface — a
+  typed nil would make the page claim a poller exists.
+- A pass that returns before computing figures (refusal backoff, sidecar down)
+  carries the previous pass's numbers forward rather than recording zeroes.
+- **`Checked` next to `Due` is what separates a stopped Lane from a quiet one**,
+  so neither may be dropped from the row.
+- Due-without-Checked is **not** by itself a stall: a browser Lane under both
+  wake thresholds sets `LaneState.Asleep` and renders "browser asleep", and
+  never counts toward `Attention`. That is the commonest healthy state for
+  kagane, comix and novelfull, so spending the stall mark on it would train the
+  owner to ignore the mark that matters.
@@ -47,6 +47,15 @@ func fetchCoverBytes(ctx context.Context, cover string, browser BrowserCoverFetc
 // inject it to exercise hostile DNS results without touching the live network.
 type CoverResolver func(context.Context, string) ([]netip.Addr, error)

+// maxCoverBytes caps one cover, separately from the series-page maxBodyBytes:
+// a cover is a bounded binary asset, not a text page, and 4 MiB rejected 12%
+// of asurascans covers measured 2026-08-17 (p90 4.52 MB, max 8.57 MB — two of
+// the three over-cap files were JPEGs, not the animated GIF of issue #71).
+// 10 MiB is ~18% headroom over that worst case and matches the GitHub and
+// Discord image limits; see docs/research/gif-maximum-byte-size.md. GIF itself
+// has no maximum size, so this number is policy, not format.
+const maxCoverBytes = 10 << 20
+
 // TLSCoverFetcher retrieves image bytes with the standard HTTPS client. Unlike
 // TLSFetcher, it does not need a browser fingerprint: cover hosts are public
 // CDNs and the response is accepted only after the destination gate passes.
@@ -152,15 +161,15 @@ func (f *TLSCoverFetcher) Fetch(ctx context.Context, sourceURL string) ([]byte,
 	if !ok {
 		return nil, "", fmt.Errorf("fetch cover: unsupported content type %q", raw)
 	}
-	if resp.ContentLength > maxBodyBytes {
-		return nil, "", fmt.Errorf("fetch cover: response exceeds %d bytes", maxBodyBytes)
+	if resp.ContentLength > maxCoverBytes {
+		return nil, "", fmt.Errorf("fetch cover: response exceeds %d bytes", maxCoverBytes)
 	}
-	body, err := io.ReadAll(io.LimitReader(resp.Body, maxBodyBytes+1))
+	body, err := io.ReadAll(io.LimitReader(resp.Body, maxCoverBytes+1))
 	if err != nil {
 		return nil, "", fmt.Errorf("read cover: %w", err)
 	}
-	if len(body) > maxBodyBytes {
-		return nil, "", fmt.Errorf("fetch cover: response exceeds %d bytes", maxBodyBytes)
+	if len(body) > maxCoverBytes {
+		return nil, "", fmt.Errorf("fetch cover: response exceeds %d bytes", maxCoverBytes)
 	}
 	return body, contentType, nil
 }
@@ -171,7 +171,7 @@ func TestCoverFetcherRejectsOversizedBody(t *testing.T) {
 	var calls int
 	client := &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
 		calls++
-		response := coverResponse(http.StatusOK, "image/webp", "", bytes.Repeat([]byte("x"), maxBodyBytes+1))
+		response := coverResponse(http.StatusOK, "image/webp", "", bytes.Repeat([]byte("x"), maxCoverBytes+1))
 		response.ContentLength = -1
 		return response, nil
 	})}
@@ -187,6 +187,29 @@ func TestCoverFetcherRejectsOversizedBody(t *testing.T) {
 	}
 }

+// Covers between the series-page cap and the cover cap must be accepted: the
+// 4 MiB page cap rejected 12% of asurascans covers (issue #71).
+func TestCoverFetcherAcceptsCoverOverPageCap(t *testing.T) {
+	body := bytes.Repeat([]byte("x"), maxBodyBytes+1)
+	client := &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+		return coverResponse(http.StatusOK, "image/gif", "", body), nil
+	})}
+	fetcher := newCoverFetcher(client, func(context.Context, string) ([]netip.Addr, error) {
+		return []netip.Addr{netip.MustParseAddr("198.51.100.10")}, nil
+	})
+
+	got, contentType, err := fetcher.Fetch(context.Background(), "https://cdn.example/big.gif")
+	if err != nil {
+		t.Fatalf("Fetch rejected a %d-byte cover: %v", len(body), err)
+	}
+	if len(got) != len(body) {
+		t.Fatalf("body = %d bytes, want %d", len(got), len(body))
+	}
+	if contentType != "image/gif" {
+		t.Fatalf("content type = %q, want image/gif", contentType)
+	}
+}
+
 func TestCoverFetcherRejectsNonImage(t *testing.T) {
 	var calls int
 	client := &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
@@ -0,0 +1,344 @@
+# GIF — maximum byte size of a file
+
+Research note for Gitea issue #71 (backend `maxBodyBytes` = 4 MiB rejects the
+8,571,192-byte animated cover GIF at
+`https://cdn.asurascans.com/asura-images/covers/a-dragonslayers-peerless-regression.gif`).
+
+All facts fetched live on **2026-08-17**: the GIF89a spec at
+`https://www.w3.org/Graphics/GIF/spec-gif89a.txt`, Go stdlib `image/gif`
+sources at `/usr/local/go/src/image/gif/reader.go` (Go 1.26.5), Chromium
+`blink/renderer/platform/image-decoders/` sources via
+`chromium.googlesource.com`, Firefox `image/decoders/nsGIFDecoder2.cpp` via
+`hg.mozilla.org`, and cover bytes probed with plain `curl` (desktop Chrome UA;
+`HEAD`/ranged `GET`). **No Cloudflare challenge was encountered on any CDN
+probe** — every request returned real headers, consistent with the AGENTS.md
+note of 2026-07-26 that plain `curl` works against both scan sites from the
+dev machine and the VPS.
+
+Every claim carries the URL it came from, or a reproducible command.
+Interpretation rather than observation is marked `[INFERENCE]`.
+
+---
+
+## 1. Summary answer table
+
+| Question | Answer | Evidence |
+|---|---|---|
+| Does the GIF89a spec define a maximum file size? | **No.** There is no file-size field anywhere in the format; the only numeric ceilings are per-field (16-bit screen/image dimensions, 255-byte sub-blocks, 12-bit LZW codes). | §2 |
+| Maximum logical screen | 65535 × 65535 pixels (unsigned 16-bit width/height). | §2.1 |
+| Number of frames / image descriptors | Unbounded — "An unlimited number of images may be present per Data Stream." | §2.2 |
+| Formal max byte size of any single GIF | None. Single-frame worst case ≈ **6.44 GB** (12-bit LZW, max canvas); animated GIFs are **unbounded** because frames are unbounded. | §3 |
+| Does the backend's decoder (Go `image/gif`) bound size? | **No.** It reads 16-bit dimensions and allocates `width×height` bytes per frame; a 65535² frame forces a ~4 GiB allocation. No total-size or dimension guard. | §4.1 |
+| Do browsers bound on-disk GIF size? | Chromium and Firefox: no on-wire size cap in their GIF readers; Chromium caps *decoded* memory at min(4 B × pixels, platform budget). | §4.3, §4.4 |
+| Real cover sizes (asurascans, n=25) | min 190,410 B · median 1,275,082 B · p90 4,524,788 B · max 8,571,192 B · **3/25 > 4 MiB** (two JPEGs and the animated GIF) | §5 |
+| Real cover sizes (demonicscans/readermc, n=78) | min 13,298 B · median 63,061 B · max 801,200 B · 0/78 > 4 MiB | §5 |
+| Comparable service caps | GitHub: 10 MB for images/GIFs. Discord API: default 10 MiB per file. Wikimedia: 100 MiB upload / 5 GiB host. | §6 |
+| Recommended cover cap for #71 | **10 MiB** (separate from the 4 MiB series-page cap). Covers 100% of the 103 observed covers; matches GitHub/Discord calibration; ≤ 20 MiB worst-case transient per concurrent fetch+serve on a 1974 MiB swapless VPS. | §7 |
+
+---
+
+## 2. What the GIF89a specification actually bounds
+
+Source: `https://www.w3.org/Graphics/GIF/spec-gif89a.txt` (fetched 2026-08-17).
+
+### 2.1 Fixed-width fields — the only hard ceilings
+
+The format is a stream of fixed-width blocks; the numeric fields that *do*
+have a ceiling are all 16-bit unsigned, little-endian ("multi-byte numeric
+fields are ordered Least Significant Byte first", §4 of the spec):
+
+- **Logical Screen Width / Height** — "Unsigned" 2-byte fields (§18, Logical
+  Screen Descriptor) → maximum **65535 × 65535** pixels.
+- **Image Left / Top Position, Image Width / Height** — "Unsigned" 2-byte
+  fields (§20, Image Descriptor). Each image "must fit within the boundaries
+  of the Logical Screen" (§20a), so an image cannot exceed the 65535² canvas
+  even though its own fields would allow it.
+- **Data sub-blocks** — "A data sub-block may contain from 0 to 255 data
+  bytes" (§15); each sub-block is preceded by a 1-byte size field and the
+  stream is terminated by a 0x00 Block Terminator (§16). This bounds a
+  *chunk*, not the stream.
+- **Global/Local Color Tables** — optional, "3 x 2^(Size of Global Color
+  Table+1)" bytes with a 3-bit size field → at most 3 × 2⁸ = **768 bytes**
+  each (§19, §21).
+- **LZW codes** — "The output codes are of variable length, starting at
+  <code size>+1 bits per code, **up to 12 bits per code**. This defines a
+  maximum code value of 4095 (0xFFF)" (Appendix F, COMPRESSION, rule 4).
+- **Trailer** — a single byte, fixed value 0x3B, "indicating the end of the
+  GIF Data Stream" (§27).
+
+### 2.2 What is unbounded
+
+- **Number of images (frames).** §20a, verbatim: "This block is REQUIRED for
+  an image. Exactly one Image Descriptor must be present per image in the
+  Data Stream. **An unlimited number of images may be present per Data
+  Stream.**"
+- **The Data Stream itself.** The grammar in Appendix B is
+  `<GIF Data Stream> ::= Header <Logical Screen> <Data>* Trailer`, and the
+  spec states "the entity Data … may be repeated any number of times,
+  including 0 times." There is **no field anywhere that carries a file size,
+  byte count, frame count, or total-length value**. §13 (Block Sizes) only
+  defines sizes *within* blocks.
+
+### 2.3 Verdict
+
+**The GIF89a specification defines no maximum file size.** The only hard
+bounds are per-field: 65535×65535 pixels per screen/image, 255 bytes per
+sub-block, 12 bits per LZW code, and one trailer byte. A compliant decoder
+must process whatever stream the blocks describe. Any byte ceiling a
+particular GIF actually hits is therefore *implicit* — 16-bit dimensions,
+LZW code width, decoder memory, or an external policy — never something the
+format itself enforces. `[INFERENCE]` This is why real-world GIFs cap out at
+"a few GB at most" and every service that wants a bound has to impose one
+itself (see §6; Wikimedia explicitly documents that a 4 GiB host limit was a
+storage-representation artifact of 32-bit integers, `phab:T191805`, not a
+format limit).
+
+---
+
+## 3. Theoretical worst case
+
+### 3.1 Single frame, maximal canvas, 8-bit pixels
+
+| Quantity | Value | Derivation |
+|---|---|---|
+| Max pixels | 4,294,836,225 | 65535 × 65535 |
+| Raw 8-bit palette-index raster | 4,294,836,225 B ≈ **4.29 GB / 4.00 GiB** | 1 byte per pixel (Table Based Image Data, §22; Go's `image.Paletted` uses exactly 1 byte/pixel) |
+| LZW worst case | ≈ **6.44 GB / 6.00 GiB** | codes ≤ 12 bits each (Appendix F), at most ~1 code per pixel for incompressible data → ≤ 12 bits/px = 1.5 B/px → 4,294,836,225 × 1.5 B |
+| Sub-block overhead | ≈ +25.3 MB | every ≤255-byte chunk carries a 1-byte size field (§15): ⌈6,442,254,338 / 255⌉ ≈ 25,263,743 size bytes, + 1 block terminator |
+| Fixed overhead | ≈ +1.6 KB | header 6 B (§17) + logical screen descriptor 7 B (§18) + global color table ≤ 768 B (§19) + image descriptor 10 B (§20) + local color table ≤ 768 B (§21) + LZW minimum code size 1 B (§22) |
+
+So a **single maximal-frame GIF cannot exceed ≈ 6.47 GB on the wire**
+(12-bit LZW bound), and LZW being lossless means the real byte count depends
+entirely on image content — the same canvas can be a few KB (flat color) or
+~6 GB (noise).
+
+Two caveats, both marked `[INFERENCE]`:
+
+- The "1.5 B/px" figure assumes ~one emitted code per pixel. An encoder is
+  permitted to emit a Clear code at any point (Appendix F: "The Clear code
+  can appear at any point in the image data stream"), so a
+  pathological-but-compliant encoder emitting clear+pixel per pixel reaches
+  ~24 bits/px ≈ 12.9 GB for the max canvas. Real encoders do not do this;
+  12-bit/px is the practical bound.
+- The spec's deferred-clear note (cover sheet) explicitly allows an encoder
+  to keep using a full table at 12-bit codes without clearing, so the 12-bit
+  cap holds for the whole stream, it cannot "grow" past 12 bits.
+
+### 3.2 Animated GIFs: unbounded
+
+Every frame is one Image Descriptor, each bounded by the 65535² canvas, but
+the *count* of frames is unbounded (§2.2). Total bytes = sum over frames —
+therefore **there is no finite maximum byte size for an animated GIF** in
+the format. The only thing that stops a real one is decoder memory, a
+service cap, or disk space. `[INFERENCE]` This is the category the issue #71
+cover falls into: it is an animated GIF (NETSCAPE2.0 loop extension found at
+offset 0x310 of the file, verified 2026-08-17 by a ranged GET), and its
+8,571,192 bytes are ~2.04× the current 4 MiB backend cap.
+
+---
+
+## 4. Decoder-side real limits
+
+### 4.1 Go `image/gif` (the backend's decoder path, stdlib)
+
+Source: `/usr/local/go/src/image/gif/reader.go`, Go 1.26.5.
+
+- Dimensions are read as little-endian uint16 — `left/top/width/height :=
+  int(d.tmp[N]) + int(d.tmp[N+1])<<8` (reader.go:490-493) — so the format
+  ceiling 65535 applies, and nothing smaller is enforced.
+- The only geometric check is that each frame fits inside the logical
+  screen: `if left+width > d.width || top+height > d.height` →
+  `errors.New("gif: frame bounds larger than image bounds")` (reader.go:512-513).
+- **There is no file-size, byte-count, frame-count, or pixel-count guard.**
+  Each frame allocates `image.NewPaletted(...)` (reader.go:515) — a
+  `[]byte` of width×height — so decoding one legal 65535² frame attempts a
+  **~4.29 GB allocation**. `DecodeAll` (reader.go:603-605) additionally
+  retains every frame's `Pix` slice for the lifetime of the returned `*GIF`.
+- `[INFERENCE]` On the 1974 MiB swapless VPS (root AGENTS.md), decoding such
+  a file would OOM rather than error cleanly; nothing in stdlib protects
+  the process. This matters for §7: the backend stores cover bytes without
+  decoding them (see §5.3), so the fetch path never triggers this — but any
+  future "validate/re-encode server-side" scheme would.
+- Grep for `MaxInt|limit|too large|bounds` in reader.go: the only hits are
+  the frame-bounds check above and the `tmp [1024]byte` scratch buffer
+  (reader.go:109); no size caps exist.
+
+### 4.2 giflib / libgif
+
+**Not verified from source.** On 2026-08-17 the giflib sources were not
+reachable from this network: `github.com/giflib/giflib` returns 404 (repo
+gone/moved), `gitlab.com/giflib/giflib/-/raw/...` answers a Cloudflare
+"Just a moment…" challenge, and the SourceForge project download path
+404s. No limit claim about giflib is made here. `[INFERENCE]` giflib is
+widely known to be allocation-driven with no dimension cap, but that is not
+checked against source and is not needed for issue #71 (the backend uses Go
+stdlib, not giflib).
+
+### 4.3 Chromium (browser behaviour, first-party source)
+
+- `third_party/blink/renderer/platform/image-decoders/gif/gif_image_reader.cc`
+  (via `chromium.googlesource.com/chromium/src/+/main/...`, fetched
+  2026-08-17): **no GIF byte-size or dimension cap found** — grep for
+  `max|limit|too large|dimension|65535|overflow` matches only license text.
+- The base `ImageDecoder` caps *decoded memory*, not transfer size:
+  `CalculateMaxDecodedBytes` computes `min(4 * num_pixels, platform_max_decoded_bytes)`
+  (8 bytes/pixel for high-bit-depth), and the header comment says "Ignoring
+  this limit can cause excessive memory use or even crashes on low-memory
+  devices"
+  (`image_decoder.cc:94-117`, `image_decoder.h:545-549`). The GIF reader
+  itself is untouched by this — it is a decoded-buffer budget.
+- Practical consequence `[INFERENCE]`: a browser will happily download and
+  store a multi-GB GIF from its own cache perspective; Chromium only limits
+  what it *decodes* into pixels.
+
+### 4.4 Firefox
+
+`image/decoders/nsGIFDecoder2.cpp` (via `hg.mozilla.org/mozilla-central/
+raw-file/tip/...`, fetched 2026-08-17): **no dimension or size limit**; the
+only guards are on LZW code width (`MAX_BITS` = 12, "maximum codeword size
+of 12 bits") and the decode stack. Nothing bounds the on-disk byte size.
+
+### 4.5 Summary
+
+No mainstream decoder enforces a byte-size ceiling; they stop at the 16-bit
+dimension ceiling (Go, by construction) or at decoded-memory budgets
+(Chromium) or nowhere (Firefox). A GIF's byte size is policed only by
+*storage* policies — which is what §6 calibrates and §7 sets.
+
+---
+
+## 5. Practical distribution — what real manga covers weigh
+
+Probed **2026-08-17** with `curl -sI` (HEAD) and ranged GETs, desktop Chrome
+UA. No Cloudflare block on any request. Sample = covers *as the backend
+would fetch them* (the `og:image`/page-listed cover URL), not thumbnails we
+chose by hand.
+
+### 5.1 Exact commands
+
+```sh
+# asurascans.com — harvest cover URLs from the homepage, then HEAD each
+curl -s -A "Mozilla/5.0 (Windows NT 10.0; Win64; x64) Chrome/126.0" https://asurascans.com/ -o home.html
+grep -oE 'https://cdn\.asurascans\.com/asura-images/covers/[^"&\\< ]+\.(webp|gif|jpg|jpeg|png)' home.html \
+  | sort -u | grep -v '\-400\.' | head -25 > sample.txt        # one full-res cover per series, no -400 thumbs
+while read -r u; do curl -s -A "…Chrome/126.0" -I "$u" | tr -d '\r' \
+  | grep -iE '^content-length:'; done < sample.txt
+
+# demonicscans.org — covers live on readermc.org (ADR-0007), URLs contain spaces/UTF-8
+curl -s -A "…Chrome/126.0" https://demonicscans.org/ -o demonic.html
+grep -oE 'src="https://readermc\.org/images/thumbnails/[^"]+"' demonic.html | tr -d 'src="' > demonic.txt
+# …plus og:image from 5 manga pages (Catastrophic-Necromancer, Magic-Emperor, …)
+# each URL percent-encoded per path segment (urllib.parse.quote, safe=':/') before HEAD
+```
+
+### 5.2 asurascans — 25 full-res covers (mixed formats)
+
+Homepage fetched 200 (664,700 B). All 25 returned `200` with a real
+`Content-Length`. Distribution:
+
+| Statistic | Bytes |
+|---|---|
+| n | 25 |
+| min | 190,410 |
+| median | 1,275,082 |
+| p90 | 4,524,788 |
+| max | 8,571,192 |
+| mean | 1,943,651 |
+| **> 4 MiB (4,194,304)** | **3 (12%)** — `a-dragonslayers-peerless-regression.gif` 8,571,192 (the issue #71 cover, animated: NETSCAPE2.0 at 0x310, 550×733, 256 colors); `bad-born-blood.3008f6.webp` 4,524,788 `image/jpeg`; `ending-maker.cfbf53.webp` 4,619,303 `image/jpeg` |
+
+Notes: the CDN serves `Content-Type` by stored bytes, not by URL extension
+(the `.webp` URLs return `image/png`, `image/jpeg`, or `image/webp` — the
+sample spans all four of `png/jpeg/webp/gif`). Two of the three over-cap
+files are **not GIFs**, so the current 4 MiB cap already silently drops 12%
+of asura covers of any format. p90 itself (4.52 MB) exceeds the cap.
+
+### 5.3 demonicscans — 78 covers on readermc.org
+
+78 unique cover URLs (73 from the homepage's `/images/thumbnails/` plus 5
+`og:image` values from manga pages — demonicscans publishes the thumbnail
+file as the full cover, so that is exactly what the backend would fetch).
+**78/78 returned 200 with a real Content-Length** (spaces and UTF-8 in the
+filenames were percent-encoded per path segment; the homepage's raw HTML
+carries `â€™`-style mojibake for curly quotes, which was repaired by
+latin-1→utf-8 re-encoding before probing).
+
+| Statistic | Bytes |
+|---|---|
+| n | 78 |
+| min | 13,298 |
+| median | 63,061 |
+| p90 | 206,994 |
+| max | 801,200 |
+| mean | 110,038 |
+| > 4 MiB | 0 |
+
+### 5.4 Reading
+
+`[INFERENCE]` asurascans covers are the heavy tail (median 1.3 MB, top
+decile > 4 MiB, occasional ~5–9 MB), demonicscans covers are tiny (all
+< 0.8 MB). A cover cap must be chosen against the *asura* distribution —
+the 8.57 MB animated GIF is not a freak one-off outlier; the 90th
+percentile already crosses 4 MiB and two JPEGs sit between 4.5–4.7 MB.
+
+---
+
+## 6. Comparable documented byte caps (first-party docs only)
+
+| Service | Cap | Source (fetched 2026-08-17) |
+|---|---|---|
+| GitHub (issues/PR comments) | **10 MB for images and gifs**; 25 MB other files; 10/100 MB video | `https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/attaching-files` — "The maximum file size is: 10MB for images and gifs … 25MB for all other files" |
+| Discord (API uploads) | default **10 MiB per file**, higher with Nitro / boost tier | `https://discord.com/developers/docs/reference#uploading-files` — "The file upload size limit applies to each file in a request. The default limit is `10 MiB` for all users" (help-center article `support.discord.com/hc/en-us/articles/115002935588` exists but answered 403 from this network on the probe date, so its figures were not verified here) |
+| Wikimedia Commons | **100 MiB** upload limit; hosting up to **5 GiB**; GIF thumbnails limited to **100 megapixels**; prior 4 GiB host cap was a 32-bit storage artifact (phab:T191805) | `https://commons.wikimedia.org/wiki/Commons:Maximum_file_size` |
+| MDN | nothing — MDN documents no byte-size limit for images; browsers impose none (see §4.3–4.4) | `[INFERENCE]` from absence in the platform docs read in §4 |
+
+Calibration takeaway: two major platforms independently land on **~10 MB**
+as the ceiling for an uploadable image/GIF (GitHub exactly 10 MB, Discord
+exactly 10 MiB), with Wikimedia the outlier at 100 MiB/5 GiB because it is a
+media *archive*. A 10 MiB cover cap is therefore squarely inside industry
+normal.
+
+---
+
+## 7. Recommendation for issue #71
+
+**Raise the cover cap to 10 MiB (10,485,760 B) — as a separate constant, not
+by moving the shared one.**
+
+Why:
+
+- **Fits the measured reality.** The largest observed cover is 8,571,192 B
+  (the issue's animated GIF) = 82% of 10 MiB; 10 MiB covers **100% of the
+  103 sampled covers** and the *entire* asura distribution, including its
+  heavy tail. 4 MiB rejects 12% of asura covers (two of them plain JPEGs).
+- **Matches industry calibration** (§6): GitHub 10 MB images/GIFs, Discord
+  10 MiB default. A 10 MiB cap is a number every engineer recognizes, and
+  it leaves ~18% headroom over the current worst observed file.
+- **Costs little on the target hardware.** The backend buffers cover bytes
+  whole during fetch (`backend/internal/latest/cover.go`: `ContentLength >
+  maxBodyBytes` rejection at :155, then `io.ReadAll(io.LimitReader(…,
+  maxBodyBytes+1))` at :158) and loads the full body per `GET /covers/…`
+  (`backend/internal/api/handlers.go`, `Cover` → `w.Write(body)`). Worst
+  case per concurrent fetch **+** serve is therefore 2 × cap = 20 MiB; even
+  ten of each concurrently is ~200 MiB of a 1974 MiB swapless VPS (~10%),
+  and the browser unit (471 MiB, root AGENTS.md) is no longer on that box.
+  The 4 MiB series-page cap is *not* the issue — measured pages run
+  100 KB–1.2 MB (`backend/internal/latest/fetch.go` comment) — so keep it.
+- **The cap is a separate knob.** Today one `const maxBodyBytes = 4 << 20`
+  (`backend/internal/latest/fetch.go:17`) gates *both* series pages and
+  covers (`cover.go` references it). Raising it wholesale would loosen the
+  page-side memory guard for no benefit; a cover-specific constant (e.g.
+  `maxCoverBytes = 10 << 20`) keeps the two policies independent. The fetch
+  already double-checks `ContentLength` and the post-`LimitReader` length,
+  so a larger constant changes nothing else.
+
+Alternatives and their costs:
+
+| Option | Cost |
+|---|---|
+| Keep 4 MiB | 12% of asura covers (incl. non-GIF JPEGs) never stored — current bug, silent missing covers. |
+| 16 MiB cap | 2× headroom over the observed max for future GIFs; +60% worst-case transient memory vs 10 MiB; diverges from the GitHub/Discord 10 MB calibration. |
+| Server-side re-encode / downscale covers | Requires decoding → Go `image/gif` allocates width×height per frame with **no guard** (§4.1); a legal 65535² GIF forces a ~4.29 GB allocation on a 1974 MiB swapless box — OOM, not an error. Also mutates bytes, which the store treats as immutable/content-addressed (ADR-0007). Highest risk, no upside at this scale. |
+| No cap | Unbounded transient memory and disk; rejected outright. |
+
+Decision is the user's; on the evidence, **10 MiB for covers, 4 MiB for
+pages** is the defensible middle.
@@ -1,93 +1,103 @@
-Guidance for OpenCode (and Claude Code) working under `userscript/`. See root `AGENTS.md` for the project-wide architecture diagram, hard constraints, and design system.
+Scope: `userscript/`.

-### Userscript structure (single IIFE, `manga-bookmark.user.js`)
+Each entry names the code that holds the truth. The prose is only what the code
+cannot tell you: rationale, invariants a refactor would break, and dated
+observations about sites we don't control.

-1. **Site adapters** — one per host, `detect(location, document)` return page `type` + IDs. Identify type/IDs from **URL regex** (most stable); pull `title` from **`og:title`** (or the page heading where a site ships no og: tags), not CSS classes. **No adapter reads a cover**: the backend acquires, stores and serves every Cover from its own origin (ADR-0007), the wire's `cover` is already an address on our origin, and `apiPut` strips any `cover` off an outgoing body.
-2. **API client** — `apiGet/apiPut/apiDelete` with bearer header; `localStorage` key `bmgr:manga:cache` for instant render + offline fallback.
-3. **Progress logic** — auto-upsert `last_chapter` only when `chapterNum >= stored last_chapter_num` (re-reading old chapters must not regress progress; unparseable -> set current). Manual panel override forces any value.
-4. **Retry queue** — every write go through `pushBookmark`/`pushDelete`, so
-   failed mutation park in `localStorage` (`bmgr:manga:queue`) and replayed on
-   next navigation, reconnect, or `refresh()`. Entries are markers
-   (`{key, op, sendStatus, attempts}`), never payloads — body read from
-   cache at send time, so one entry per key give ordering and coalescing for
-   free. `sendStatus` is **sticky**: while archive pending, later writes to
-   that key keep carrying bucket, which stop successful
-   in-between write from silently un-archiving series. `refresh()` drains
-   before it fetches and overlays anything still pending, so list never
-   flaps. 400 drops entry, 401 abort pass and keep queue, and
-   transient failures retry to cap of 10. Latest-chapter writes deliberately
-   stay out of queue. See
-   `docs/superpowers/specs/2026-07-27-offline-retry-queue-design.md`.
-5. **UI** — rendered inside **Shadow DOM** root to isolate from site CSS
-   (critical on mobile). Three tabs (All / Favourites / Archived) and row of
-   link chips to web UI and both manga sites; `WEB_BASE` sits in CONFIG
-   block next to `API_BASE`. FAB is `7 × 44` edge tab whose *hit* area
-   widened to `28 × 72` by invisible `#hit` child; `#fab` must keep
-   `touch-action: none` and must **not** regain `overflow: hidden`. Since
-   `touch-action` resolved at gesture start, strip can't be both
-   browser-scrolled and script-dragged, so `makeDraggable` splits by intent: swipe
-   from `#hit` scrolls via `window.scrollBy`, hold of `ARM_MS` arms
-   reposition drag, visible sliver drags with no hold. See
-   `docs/superpowers/specs/2026-07-28-edge-tab-hitbox-design.md`.
-6. **SPA navigation** — Asura is Astro, client-routed on comic/chapter pages: patch `history.pushState`/`replaceState` + listen `popstate`, re-run `detect()` on URL change so auto-update fire without reload. Demonic uses classic reloads (initial `document-idle` run suffice).
+### Structure — single IIFE, `manga-bookmark.user.js`

-### Live URL shapes (verified 2026-07-26, may drift — re-check against live pages before trust)
+Six parts, in file order: site adapters, API client, progress logic, retry
+queue, UI, SPA navigation.

- **asurascans.com**: series `/comics/<slug>` (slug carries trailing
-  site-wide build-hash suffix, e.g. `-059befe1`, that **rotates on every
-  redeploy**), chapter `/comics/<slug>/chapter/<n>`. `seriesId` must strip
-  hash (`/-[0-9a-f]{8}$/`, `stripBuildHash` in userscript,
-  `asuraBuildHash` in backend); URLs keep full slug — stale-hash
-  URLs 302 to current ones. Astro-rendered; chapter links present in raw
-  server HTML.
- **demonicscans.org**: series `/manga/<slug>` (slug may URL-encode punctuation, e.g. `%2527` for `'`), chapter `/title/<slug>/chapter/<n>/<page>` (older `chaptered.php?manga=<id>&chapter=<n>` form still exists as redirect, what series-page chapter-list anchors link through).
-  Encodings (incl. triple-encoded punctuation like `%25252D`) identical
-  on /manga/ and /title/ pages, so decode-once seriesIds match — verified
-  2026-07-28.
- **comix.to**: series `/title/<id>-<slug>`, chapter
-  `/title/<id>-<slug>/<uploadId>-chapter-<n>`. Only the leading `<id>` is
-  identity — the slug re-renders when a series is renamed (`comixSeriesId`).
-  An SPA that **never rewrites `og:title`**: the server-rendered head keeps
-  whatever document loaded first, so on a cold load `og:title` is the homepage's
-  "Comix — Read Comics online for free" and after an in-page hop it is the
-  *previous* series' name. `document.title` is the one thing client routing does
-  update, so titles come from there, with the chapter page's `" · Ch.<n>"` tail
-  stripped. It publishes no `og:image` either, which is one of the reasons cover
+**Site adapters** — one per host, `detect(location, document)` returning page
+`type` + IDs.
+
+- Identify type and IDs from **URL regex**, which is the most stable surface a
+  site exposes; take `title` from **`og:title`** (or the page heading where a
+  site ships no og: tags), never CSS classes.
+- **No adapter reads a cover.** The backend acquires, stores and serves every
+  Cover from its own origin, the wire `cover` is already an address there, and
+  `apiPut` strips any `cover` off an outgoing body.
+
+**Progress logic** — auto-upsert `last_chapter` only when
+`chapterNum >= stored last_chapter_num`; unparseable sets the current value.
+Re-reading an old chapter must not regress progress. A manual panel override
+forces any value.
+
+**Retry queue** — every write goes through `pushBookmark`/`pushDelete`.
+
+- Entries are markers (`{key, op, sendStatus, attempts}`), **never payloads**:
+  the body is read from cache at send time, so one entry per key gives ordering
+  and coalescing for free.
+- `sendStatus` is **sticky** — while an archive is pending, later writes to that
+  key keep carrying the bucket. Without it a successful in-between write
+  silently un-archives the series.
+- `refresh()` drains before it fetches and overlays anything still pending, so
+  the list never flaps.
+- 400 drops the entry, 401 aborts the pass and keeps the queue, transient
+  failures retry to a cap. Latest-chapter writes deliberately stay out of the
+  queue.
+
+**UI** — rendered inside a **Shadow DOM** root to isolate it from site CSS,
+which is critical on mobile.
+
+- The FAB's *hit* area is widened by an invisible `#hit` child. `#fab` must keep
+  `touch-action: none` and must **not** regain `overflow: hidden`.
+- `touch-action` is resolved at gesture start, so the strip cannot be both
+  browser-scrolled and script-dragged. `makeDraggable` therefore splits by
+  intent: a swipe from `#hit` scrolls via `window.scrollBy`, a hold of `ARM_MS`
+  arms a reposition drag, and the visible sliver drags with no hold.
+
+**SPA navigation** — Asura is Astro and client-routes on comic/chapter pages, so
+`history.pushState`/`replaceState` are patched and `popstate` listened to, and
+`detect()` re-runs on URL change. Demonic uses classic reloads, where the
+initial `document-idle` run suffices.
+
+### Live URL shapes
+
+Encoded in the adapters; the notes below are the parts a reader of the regex
+would get wrong. **Verified 2026-07-26 unless dated otherwise — sites drift, so
+re-check against a live page before trusting any of it.**
+
+- **asurascans.com** — the series slug carries a site-wide build-hash suffix
+  (e.g. `-059befe1`) that **rotates on every redeploy**, so `seriesId` must
+  strip it (`stripBuildHash` here, `asuraBuildHash` in the backend) while URLs
+  keep the full slug — stale-hash URLs 302 to current ones.
+- **demonicscans.org** — slugs may URL-encode punctuation, and the older
+  `chaptered.php?manga=<id>&chapter=<n>` form still exists as a redirect, which
+  is what series-page chapter-list anchors link through. Encodings (including
+  triple-encoded punctuation like `%25252D`) are identical on `/manga/` and
+  `/title/` pages, so decode-once seriesIds match (verified 2026-07-28).
+- **comix.to** — only the leading `<id>` is identity; the slug re-renders when a
+  series is renamed (`comixSeriesId`). It is an SPA that **never rewrites
+  `og:title`**: the server-rendered head keeps whatever document loaded first,
+  so on a cold load `og:title` is the homepage's name and after an in-page hop
+  it is the *previous* series'. `document.title` is the one thing client routing
+  updates, hence titles come from there with the chapter page's `" · Ch.<n>"`
+  tail stripped. It publishes no `og:image` either, one of the reasons cover
  acquisition moved to the backend.
- **kagane.to**: series `/series/<uuid>`, reader
-  `/series/<uuid>/reader/<bookUuid>`. Reader URLs carry no chapter number, so
-  the number comes out of `og:title`. Two shapes exist: `"<Series> - Chapter
-  <n>[ - Episode <n>]"` and, for volume-numbered series, `"<Series> - Volume <v>
-  Chapter <n>"` with no episode name — both must yield a bare series title, or
-  the volume tail lands in the bookmark's title.
-  Its covers are challenge- and CORP-protected, so nothing outside kagane.to can
-  load one directly; the panel renders the backend's own cover address like every
-  other Site. Behind a Cloudflare JS challenge, so the backend polls it
-  through the headless browser.
- **novelfull.com** (novel script): series `/<slug>.html`, chapter
-  `/<slug>/chapter-<n>[-<title-slug>].html`. No `og:*` tags at all — title from
-  `h3.title` (series) or `a.truyen-title` (chapter); the script reads no cover.
-  Behind a Cloudflare JS challenge no TLS fingerprint
-  clears, so the backend polls it through the headless browser.
- **lightnovelworld.net** (novel script): series `/novel/<slug>/`, chapter
-  `/<slug>-chapter-<n>/` — flat, at the site root. The chapter path's slug is a
-  Chapter Slug, not an identity: the Series address is read off the page's
+- **kagane.to** — reader URLs carry no chapter number, so the number comes out
+  of `og:title`. Two shapes exist, `"<Series> - Chapter <n>[ - Episode <n>]"`
+  and `"<Series> - Volume <v> Chapter <n>"`; both must yield a bare series
+  title, or the volume tail lands in the bookmark's title. Its covers are
+  challenge- and CORP-protected, so nothing outside kagane.to can load one —
+  the panel renders the backend's cover address like every other Site.
+- **novelfull.com** (novel script) — no `og:*` tags at all, so the title comes
+  from `h3.title` (series) or `a.truyen-title` (chapter).
+- **lightnovelworld.net** (novel script) — chapter paths are flat at the site
+  root and their slug is a **Chapter Slug, not an identity**: a Series may
+  publish under several. The Series address is read off the page's
  `a[aria-label='All Chapter']` (fallback: the BreadcrumbList's second crumb),
-  and a Series may publish under several Chapter Slugs. A chapter page with no
-  pointer resolves to `other`, so no Bookmark is offered. `h1.entry-title` is
-  the clean title on a series page and `<Title> Chapter <n>` on a chapter page.
-  Its series page lists every chapter with an
-  absolute href, so the backend polls it with the plain TLS client.
-  The client performs no latest-chapter scan for this Site: the Poll's
-  one-hour cooldown dominates the client's four-hour throttle, so a scan
-  would add no freshness, and the page's wpdiscuz thread is a public write
-  surface a scan would have to truncate at. `computeLatestChapter` yields
-  null here and `backgroundRefreshLatest` skips the Site before any fetch.
+  and a chapter page with no pointer resolves to `other` so no Bookmark is
+  offered. The client runs **no latest-chapter scan** for this Site —
+  `computeLatestChapter` yields null and `backgroundRefreshLatest` skips it
+  before any fetch — because the backend Poll's one-hour cooldown dominates the
+  client's four-hour throttle, so a scan would add no freshness while having to
+  truncate at the page's wpdiscuz thread, a public write surface.

-### Second script: `novel-bookmark.user.js`
+### Second script — `novel-bookmark.user.js`

 A copy of the manga script with two adapters, `LIBRARY = "novel"` and
-`STORE_PREFIX = "bmgr:novel:"`. No migration loop (this script has no previous
-installation to carry keys over from). Installed alongside the manga script;
-both write to the same backend with the same `LIBRARY` column discriminating
-them.
+`STORE_PREFIX = "bmgr:novel:"`. No migration loop, because this script has no
+previous installation to carry keys over from. Installed alongside the manga
+script; both write to the same backend, discriminated by `LIBRARY`.
Author	SHA1	Message	Date
sulthan	2b96b43fd0	Add a gitea skill; point AGENTS.md at it Forge usage was spread across AGENTS.md, docs/agents/issue-tracker.md, and whatever the agent remembered about `gh`. Move the how-to-run-`tea` half into a model-invoked skill so it loads on any issue/PR task, and leave AGENTS.md a single line of identity plus pointers. The skill carries only what `tea --help` cannot answer: `gh` is absent so the harness's `issue://`/`pr://` URIs error out, issues and PRs share one index space, rendered-box output needs `-o json`, `close` takes no `--comment`, labels never auto-create, bodies need a heredoc, and a read without `--comments` silently drops every comment (issue #123: 40 lines bare, 132 with the flag). Tracker conventions and triage label strings stay in docs/agents/, cited rather than copied.	2026-08-17 17:50:09 +07:00
sulthan	766aa8f00d	docs: make every AGENTS.md cite code, not docs or issues (#113 ) Every `AGENTS.md` now cites code and nothing else. ## Why Two rot mechanisms, same symptom — an agent confidently follows a stale statement: 1. Non-code citations. A spec, ADR, plan file, or issue records what was true when it was written. Nothing updates it when the decision reverses. 2. Prose restating mechanism. The code changes, the paragraph doesn't, and the next reader trusts the paragraph. Code is the only source true at read time. ## What changed All three files: removed every ADR ref, spec/plan pointer (`docs/superpowers/specs/`, `plans/`, `docs/research/`), `DEPLOY.md`/`REDEPLOY.md`, `docs/agents/`, and issue number. Facts those links carried are restated inline — the `tea` command set and the five triage label strings now live in the root Forge section. `### Domain docs` is deleted: it pointed only at `CONTEXT.md` and `docs/adr/`, neither of which exists. `backend/` and `userscript/`: rewritten around derivability. \| Class \| In code? \| Treatment \| \|---\|---\|---\| \| Structure — packages, routes, env vars, columns \| yes \| name the symbol, nothing else \| \| Mechanism — what a function does \| yes \| symbol + one line \| \| Rationale — why, what a "simplify" breaks \| no \| written out \| \| Measurement — observation against a service we don't control \| no \| written out, dated \| `backend/AGENTS.md` 20578 → 15512 bytes, `userscript/AGENTS.md` 7129 → 5912. Root grows 16905 → 19292: the cost of inlining the `docs/agents/` facts plus the new rule. Rule* recorded in root as `## Writing an AGENTS.md`. Sole non-code exception is a sibling `AGENTS.md`. Closing clause: every symbol named must exist, since a dead pointer is a bug rather than a stale sentence. Harness-agnostic: dropped the `Guidance for OpenCode (and Claude Code)` openers for plain scope lines. ## Verification Applied the new rule to itself — extracted all 118 backticked identifiers across the three files and checked each against every `.go`, `.js`, `.sql`, `.html` and `.css` source. Zero repo symbols missing; the 8 non-matches are external (`GM_setValue`, `navigator.webdriver`, `HeadlessChrome`, `curl`, …). That check caught a claim that was already lying on `main`: the cover section said `CoverFetcher` was gone, but `NewCoverFetcher`, `TLSCoverFetcher` and `BrowserCoverFetcher` are all live in `internal/latest`. Now names only the genuinely dead `/img/kagane/{id}` route. Exactly the failure the rule exists to prevent. No code touched — documentation only, nothing to test. Reviewed-on: #113 Co-authored-by: Sulthan Zaki <sultankiki05@gmail.com> Co-committed-by: Sulthan Zaki <sultankiki05@gmail.com>	2026-08-17 13:43:49 +07:00
sulthan	550b258c59	fix: give covers their own 10 MiB byte cap (#71 ) (#112 ) Co-authored-by: Sulthan Zaki <sultankiki05@gmail.com> Co-committed-by: Sulthan Zaki <sultankiki05@gmail.com>	2026-08-17 12:06:14 +07:00
sulthan	3ac865cd08	chore: remove graphify (#111 ) Removes the graphify integration. It was measured against this repo rather than assumed. ## Why `graphify query` returns a keyword-seeded BFS neighbourhood, not a location. Asked where CORS origin reflection is implemented, it returned 73 nodes — mostly `api_test.go` helpers, plus a `Reflection and Type Assertions` section from `.agents/skills/golang-performance/references/cpu.md` matched on the word "reflection" — and never named `httpmw/middleware.go:135` or `main.go:121`. `grep` returned both in 39ms. Same shape asking how the poller skips kagane: 145 nodes, top hits `poller_test.go` helpers and two nodes named `T`. `graphify explain "BrowserFetcher"` is sound (`browser.go L52`, 9 `EXTRACTED` edges), but that is what `lsp references` already answers, against live files instead of a snapshot. Staleness was never the problem — `graph.json` rebuilt 5s after `f568fb5`, so the git hooks worked. Retrieval quality was. ## What it cost - Two `PreToolUse` hooks injecting a "MANDATORY: run graphify query first" paragraph into context on every grep/find and every source-file read. - 685k input tokens across 5 build runs (`cost.json`). - 3.4MB of `graph.json` + `graph.html` tracked, across 11 commits of map-refresh churn. `AGENTS.md` is the stronger orientation artifact for a repo this size: it carries the CDP constraints, the UTC-clock finding, the per-site adapter list, and the security invariants — none of which an AST graph derives. Graphify earns its keep on repos too large to grep coherently and without curated docs; not this one. ## Changes - Delete the committed map (`graphify-out/`, -58k lines). - Drop the `## graphify` rules block from `AGENTS.md` (`CLAUDE.md` is a symlink, so both). - Drop the five `graphify-out/` entries from `.gitignore`. - Empty the two `PreToolUse` hooks in `.claude/settings.json`. - Remove the stale `graphify query` instruction from `.claude/skills/implement-tickets/SKILL.md` — it pointed dispatched ticket-implementer agents at a binary that no longer exists. Uninstalled outside the tree (not in this diff): the `graphifyy` CLI, `~/.claude/skills/graphify/`, the global `~/.claude/CLAUDE.md` block, the `Bash(graphify query )` permission in the git-ignored `.claude/settings.local.json`, and the `post-commit` / `post-checkout` git hooks. ## Verification `grep -ri graphify` over the worktree is clean; remaining hits are inside `.git/` (commit messages, two stale branch configs). No code touched — backend and userscript are untouched, so `go test ./...` is unaffected. Reviewed-on: #111 Co-authored-by: Sulthan Zaki <sultankiki05@gmail.com> Co-committed-by: Sulthan Zaki <sultankiki05@gmail.com>	2026-08-17 11:43:15 +07:00