fix(host): a share-seq reservation is not a line the room said (#2855) - #2863

Fixes #2855

Owner: Dev-3
Reviewer: Birfy
Written-by: Dev-3
Bot: Dev-3
Claim-Id: 1c8863ae0c9846ea917e3eb9dcd99282
Risk tier: T2
Base: bdaff4f
Tip: 72276f3
Tree: 4fa3f6add7fe2cf1de7d02699ae1492d86a39b42

What changed

A regression #2839 introduced. That fix made read_group_chat count a trimmed log honestly: the total is kept.length plus room.seq - room.log.length, which reads "every seq number above the retained window was a line the cap dropped". takeLineSeq broke that premise: it reserved a remote line number for a co-seat mention by moving room.seq and wrote no line, so each reservation looked like a dropped line and the header grew one phantom row per mention — a five-line log said newest 5 of 6, oldest first after one reservation and still showed all five.

The reservation now keeps its own counter, room.shareSeq, which starts above everything already said and is never room.seq; room.seq keeps its meaning (one per written line), so the trimmed-room total is right again. #1126 is unchanged in what it needs — a second mention still gets a new number, so its envelope is a new remote line instead of a redelivery of seq 0. docs/subsystems/core-host.md records the reservation rule and docs/subsystems/plugin-runtime.md (owner of hooks.js) corrects the takeLineSeq sentence.

Red first (base bdaff4f, test-only tree)

core/test/room-read-trimmed-count.test.js — five lines, one takeLineSeq reservation, then the same read:

not ok 6 - \#2855: a share-seq reservation writes no line, so the read is byte-identical
  message: 'a reservation writes no line, so nothing was left out'
  + 'Latest in "Ops" (newest 5 of 6, oldest first):\n' +
  - 'Latest in "Ops" (oldest first):\n' +
      'Writer: message 1: the merge queue slipped again\n' +
      ...
# tests 6
# pass 5
# fail 1

Two runs, normalized only for duration_ms and the random /tmp/ds-bot-p684-* home directory, are byte-identical (cmp clean).

Ran

(cd core && DSH_HARNESS=… npm test)                     # 3090/3093; the 3 red are red on the base too
(cd core && DSH_HARNESS=… node --import ./test/harness-resolve.mjs --test test/room-read-trimmed-count.test.js)   # 6/6
(cd core && DSH_HARNESS=… node --import ./test/harness-resolve.mjs --test room-read-*, room-share-braces, room-collaboration, room-context, groups, group-work, room-membership, remote-room)  # 225/225
node scripts/ds-bot-patches.mjs generate && node scripts/ds-bot-patches.mjs check     # ok, 10 patches, core/ equals ds-bot/ + patches
node core/scripts/build-client.mjs --check                                           # exit 0
ARCH_DOCS_BASE=bdaff4fb92cb… node core/scripts/check-arch-docs.mjs                   # OK (7 changed files)
node .github/scripts/changelog.mjs check                                             # 15 fragments well formed
(cd core && DSH_HARNESS=… npm run lint)                                              # 0 errors (88 warnings, pre-existing)

The 3 red tests in the full suite are red on base bdaff4fb by the same commands: bench-preflight refuse/install (pnpm absent on this box) and thinking-row #1718 ('min(560px, 78%)' vs 'min(560px,78%)'); 16/19 pass in those two files on the base.

Not run

Plugin suites (no plugins/** change) and the root npm run verify; smoke bench (no index.js/src/host/** smoke surface — hooks.js only, but see Ran for the host suite).

Generated

generate rewrote patches/ds-bot/files/src/host/hooks.js and patches/ds-bot/files/test/room-read-trimmed-count.test.js (both are whole-file mirrors; hooks.js is a FlyBot-added file, so no *.patch changed). patches check reports core/ equals ds-bot/ + patches/ds-bot/.

Risk tier

risk-tier: T2; basis: core/src/host/hooks.js=T2, changelog.d/2855-room-read-take-line-seq.fixed.md=T1, core/test/room-read-trimmed-count.test.js=T1, patches/ds-bot/files/src/host/hooks.js=T1, patches/ds-bot/files/test/room-read-trimmed-count.test.js=T1, docs/subsystems/core-host.md=T0, docs/subsystems/plugin-runtime.md=T0

T2 because of the core/src/host/hooks.js path, computed by the base copy of risk-gate.mjs and .github/agent/risk-rules.json; no gate file changed, so no Policy-sign: line.

Checkpoint

{"pr":2863,"claim":"Dev-3","issue":2855,"summary":"read_group_chat counted a share-seq reservation as a line the room said, so a co-seat mention that wrote no line printed one phantom dropped line and the count grew per mention. takeLineSeq now keeps room.shareSeq, starting above everything already said, and never moves room.seq, which still counts written lines only. A five-line log therefore reads byte-identically before and after a reservation, a second mention still gets a new number, and a room whose log really was trimmed still reports what the cap dropped. Tests pin both on the real runtime.","files":["core/src/host/hooks.js","core/test/room-read-trimmed-count.test.js","docs/subsystems/core-host.md","docs/subsystems/plugin-runtime.md","changelog.d/2855-room-read-take-line-seq.fixed.md","patches/ds-bot/files/src/host/hooks.js","patches/ds-bot/files/test/room-read-trimmed-count.test.js"],"checks":{"ran":"node --import ./test/harness-resolve.mjs --test test/room-read-trimmed-count.test.js","result":"green","at":"72276f3ef224fbea0de1fa3df5fe4f5956f49629"},"next":"none — ready for review"}