§5.5 stale-reply withholding: thread-follow replies yield to a room that moved mid-turn

[Octember]

Why

Live 2026-07-23 ("bevelina is back to being an idiot"): in the font/image incident thread, Nina asked Noah a question; Noah answered it himself 14 seconds into Bevelina's thread-follow wake; her reply — composed against the wake-start snapshot — landed a minute later, stale and partly wrong, and the following wake misread Noah's answer as a correction of her and posted a flip-flop. Not rot (M20 healthy), not model routing — the SPEC §5.5 staleness guard was unimplemented.

What

Implements SPEC §5.5 / the §18.2 "Stale-reply withholding" row:

• In a wake whose batch contains no direct address, replies buffer until turn end. If newer addressed messages arrived on the same conversation mid-turn, the reply is withheld (ledger effect kind: withheld, never a false posted) and the immediately following wake carries it as an unsent draft to reconsider — the model's words, the model's call. Otherwise it posts normally at turn end (home thread still via the reply stream).
• A directly-addressed wake never buffers: the asker is owed the answer even if the thread moved (SPEC MUST NOT).
• toolset: optional bufferReply seam on ToolsetContext; a buffered reply records no effect at call time.
• turn: beforeRecord hook so the flush lands before the turn row records — the ledger stays truthful.
• service: buffer + flush; failed/timed-out attempts drop their buffer (same rule as clearCards); unsent drafts held in-memory like ear notes (fail-open).

Tests

Three new §18.2 conformance rows in test/resident.test.ts (withheld + draft carried; no-arrivals posts at turn end; direct never withheld). bun run check green: 361 pass.

Deployed to tag-daemon.exe.xyz 2026-07-23 12:50 UTC (VM base verified against parent commit before scp).

🤖 Generated with Claude Code

[Octember]

Second commit on this branch (e0d39b6): the over-helpfulness half of the regression. The ear was waking her for conversations it itself judged to belong to someone else ("noah owns the incident response", "kate still owes that check", "someone should investigate") and recording third-party debts that rode her wake prompt as hers — ~40 threads/day posted into on Jul 21–22. Fix: one ownership boundary in the ear soul (a question aimed at another teammate is theirs even when she knows the answer; record only debts aimed at her) + matching verdict-tool description. Alert front-running stays licensed via her memory facts. Deployed to the VM 2026-07-23 ~14:05 UTC; the one open third-party attention item was operator-closed.

[Octember]

Third commit (2481ae3), from the "stfu youre not helping" evidence: the ear was converting open team asks into her personal debts — "noah is asking you to QA BEV-4611" was recorded seconds after Noah said "not for u" in the same channel, and "the gtm team, including you" invented her team membership, producing the "i've got capacity" interjection. Both QA debts were recorded anchor-less (thread_root_id NULL) so her step_backs could never settle them, and the ear reopened one three times on work-still-outstanding logic. Fixes: open_ask now requires an anchor (top-level asks root on their own ts), operator closes are ear-irreversible, and the ear soul gets the open-ask boundary (open work is not hers to claim unless a name or standing rule makes it hers; unfinished work is not an unanswered ask). Deployed ~15:20 UTC; the two orphaned QA debts operator-closed; her open-debts list is now empty.