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>

AGENTS.md

@@ -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 +67,29 @@ 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:
+`origin` is self-hosted Gitea instance (`gitea.violetcrown.my.id`, repo `sulthan/mangaBookmark`), so **`gh` don't work here — use `tea` (Gitea CLI) for anything past plain git.** `tea` infers the repo from `origin`; auth lives in `tea login`, not a `GH_TOKEN` env var. It prints rendered boxes rather than plain text, so pass `-o json` when parsing; a PR URL lands on the last line.
 
-- 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.
+- PRs: `tea pr create --head <branch> --base main --title "..." --description "..."`, `tea pr list`, `tea pr <n>`, `tea pr checkout <n>`.
+- Issues: `tea issue create --title "..." --description "..."` (`--labels`, `--assignees` optional), `tea issue <n> --comments`, `tea issue list --state open|closed|all -o json`, `tea issue close <n>`.
+- Comments: `tea comment <n> "..."` — `tea issue close` takes no `--comment` flag.
+- Labels: `tea issue edit <n> --add-labels "..."` / `--remove-labels "..."`. Gitea will **not** auto-create a label, so `tea labels create --name "..." --color "#rrggbb"` first.
+- Triage vocabulary is `needs-triage`, `needs-info`, `ready-for-agent`, `ready-for-human`, `wontfix`.
+- Gitea shares one index space across issues and PRs, so a bare `#42` may be either — try `tea pr 42`, fall back to `tea issue 42`.
 
 ## 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 +131,19 @@ 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 +152,49 @@ 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.