feat: one Poll Lane per Site, replacing the shared pace (#100)

Each Site runs its own poll goroutine, paced by Rest and Gap from the
registry (sites.go) instead of the shared cooldown/interval/stagger/batch
config. Rest is enforced by the due query's WHERE clause; the Lane sleeps
its effective gap between fetches (hour/eligible, floored at 1s).

Lane-local failures: two refusals stop that Site for 15m, a lost browser
stops only the round's remaining browser Lanes, and cover heals moved to
background goroutines so a slow CDN cannot consume a Lane's gap. The five
LATEST_CHAPTER_POLL_* pace env vars are gone; only the kill switch
remains.

docs/adr/0010-poll-lanes-per-site-pace.md

@@ -0,0 +1,91 @@
+# ADR-0010: Poll Lanes — one independent Poll stream per Site
+
+Date: 2026-08-16
+Status: accepted
+
+## Decision
+
+Replace the single shared polling pace with one **Poll Lane** per Site: an
+independent goroutine that polls only that Site's Series, paced by that Site's
+registry entry. Pace moves out of config and into the Site registry
+(`internal/latest/sites.go`): every entry carries a `Rest` (how long a Series
+rests between Polls) and a `Gap` (how long the Lane waits between fetches).
+
+Rest is enforced by the due query's WHERE clause (`latest_checked_at <= now -
+Rest`), never by a timer — the same mechanism that enforced the old cooldown.
+The Lane enforces its own gap by sleeping between fetches. `effectiveGap` is
+the registry gap, or one hour divided by the Site's eligible Series count when
+that is smaller, never below one second.
+
+The five environment settings that used to size the shared pace —
+`LATEST_CHAPTER_POLL_COOLDOWN`, `_BROWSER_COOLDOWN`, `_INTERVAL`, `_BATCH`,
+`_STAGGER` — are deleted. Only the kill switch `LATEST_CHAPTER_POLL_ENABLED`
+remains. No deployed `.env` may carry the deleted knobs.
+
+## Why
+
+The shared pace capped the whole backend at roughly 180 Polls an hour (one
+20-second stagger across one queue). ~60 Series today, scaling to hundreds or
+thousands, would stretch the hour beyond what the New Chapter signal can
+tolerate. Worse, the queue mixed Sites with very different costs: kagane and
+comix pay seconds of a serialized single-tab Chrome per Poll (a challenged
+page, ADR-0005), and one hostile Site burning its challenge timeout made every
+other Site's Series wait — "one hostile Site can eat most of an hour".
+
+Lanes fix both at once:
+
+- **Throughput scales per Site.** The six Lanes fetch concurrently; a Lane's
+  own gap paces it. The browser Lanes' combined ceiling stays about 360 Polls
+  an hour (one tab), and when they cannot keep up the wait past Rest grows and
+  is logged every pass — the "behind by X" measurement, so the decision to
+  give browser Sites more pages is made from data.
+- **Hostility is contained.** A refusal (two challenge-held reads in one
+  pass) stops only that Site's Lane for `refuseBackoff` (15m); the rest of
+  that Lane's Series stay unstamped and due. A lost browser gates the other
+  browser Lanes' passes for the same window — the flag is shared Poller
+  state, so the loss is noticed once instead of once per Lane per pass, and
+  decays after 15m so the Lanes probe again. One Site can no longer tax the
+  others.
+
+## Tradeoffs and rejections
+
+- **Per-Site env knobs** (e.g. `KAGANE_POLL_GAP`) rejected: the registry is
+  the single place pace lives, testable and reviewable; config knobs would
+  recreate the shared-pace sprawl with six times the surface. All six entries
+  are deliberately uniform at first — rest an hour, gap ten seconds — so the
+  structure exists to differ without inventing numbers for Sites that have
+  not earned them.
+- **Dynamic gap** (`rest / eligible`) is the one knob that stays automatic:
+  a Site with more Series than one per ten seconds would otherwise back up
+  behind its own gap, and the per-Series share of the hour is the natural
+  pace. The ten-second default is not arbitrary: one request per ten seconds
+  is the strictest rate rule a free-plan Site can even express (per-zone
+  rate limiting, as documented in
+  `docs/research/cloudflare-bot-scoring-and-poll-cadence.md`), so the
+  default pace is exactly what the most restrictive Site would demand of us.
+  The computed gap never goes below one second and logs loudly when the
+  floor engages.
+- **Timer-based pacing** rejected: the old ticker made the poller's rate a
+  function of wall clock rather than of what was due. The due-query cutoff is
+  the only rate authority; the Lane sleep just prevents hammering.
+- **Batch size** (the old `_BATCH` cap) is gone with the shared pace: a Lane
+  processes everything due, paced by its gap. There is no global queue left
+  to bound.
+
+## Constraints preserved
+
+- Stamp-before-fetch ("attempted" semantics): an untried Series stays due, so
+  a browser that appears after a restart finds its full queue waiting.
+- Browser wake gate (ADR-0005): a browser Lane leaves Chrome asleep below
+  five due Series and 15 minutes of wait, per Lane.
+- The browser is not in the API stack (ADR-0006): an unreachable browser
+  degrades a Lane exactly as an unset `BROWSER_WS_URL` — browser-only Sites
+  skipped, plain-TLS unaffected, stored covers still served.
+- Cover heals moved to background goroutines (joined by the test suite via
+  `waitCovers`) so a slow cover CDN cannot consume a Lane's gap.
+
+Supersedes the pace mechanics of ADR-0003's "raise throughput instead" note
+(the stagger cut it rejected is what the per-Lane gap replaces) and the
+6-hour browser cooldown introduced with the browser-backed Sites; the
+1-hour browser rest was already cleared as safe by
+`docs/research/cloudflare-bot-scoring-and-poll-cadence.md`.