Seat-post protocol: registry bodies grow without bound — bound the live body to current state, move shift narration out #7583

Problem

Seat posts (label:pm:seat) are defined as "the single authoritative registry" for each seat, but nothing bounds their size. #6019 (domain:engine-core) had reached ~61 KB by 2026-08-11: wave-by-wave status narration, per-card verdicts, patrol notes and stacked handover sections all accreted in the body across shifts. The maintainer raised it directly today (2026-08-11, engine-core PM session):

> 座位贴 #6019 内容较大, 这个问题应该怎么解决，以后会越来越大。

Cost is concrete, not aesthetic: a 61 KB body exceeds tool read limits (this morning's takeover needed chunked shell reads of a saved file to read its own seat post), every full-body edit must round-trip the whole text through a write channel, and the operative state — which the protocol says the body exists to carry — is buried under history that belongs to closed cards.

Growth mechanism

The protocol already says the right things — "评论只作交接审计,不承载状态", state lives on the issues — but in practice the body absorbed shift narration ("#NNNN DISPATCHED…", "#NNNN ACCEPTED + QUEUED…", per-wave tallies), which is per-card state that already lives on the cards and PRs. That narration is the unbounded term; the genuinely body-shaped sections (Scope / Current PM / inherited ledger / hot-file serial queue / standing commitments) are all naturally bounded.

Precedents already in the field

• [PM seat] domain:engine — 🟢 os-elon (session_019yDEhPBC3tcGkW9bkce1HM)(合并后车道:engine-core+metadata+drivers) #6019 was fully re-authored once before (Chinese → English rewrite) relying on body edit history as the archive — established that edit history is an acceptable archive channel.
• The 2026-08-11 06:0xZ takeover of [PM seat] domain:engine — 🟢 os-elon (session_019yDEhPBC3tcGkW9bkce1HM)(合并后车道:engine-core+metadata+drivers) #6019 compacted the live body to current-state-only (~4.5 KB), with prior logs retained in the 05:13Z body revision. Worked; nothing was lost; the takeover audit comment records the pointer.

Proposal (for a SKILL.md amendment — ⚠️ any card editing .claude/skills/pm-dispatch/** mandates model: claude-fable-5 per the 2026-08-10 Model-tiering ruling)

1. Fixed section template for seat bodies: Scope / Current PM / Inherited state / Hot-file serial queue / Standing commitments / Notes — nothing else. All sections are current-state; none grow with time.
2. Shift narration goes where the state already lives: per-card dispatch/accept/landing facts belong on the card and PR threads (they are already required there); round tallies go to the chat round report. The seat body is edited at takeover, handover, and structural change — not per event.
3. Takeover compaction as a standard step: an incoming PM restates the handover's operative content in the bounded template and points at the outgoing revision in edit history (audit comment carries the pointer).
4. Mechanical guard: a soft size predicate (e.g. body > ~10 KB) added to scripts/pm/check-half-states.mjs as a report-only H-item, so oversized seat posts surface in patrol sweeps instead of being discovered at read-failure time.

Options 1–3 are protocol text; option 4 is a small script change. The #7341-style skill-batch card is the natural vehicle if one is already queued.

Filed unassigned, no domain:* — triage routes (likely domain:devx by the skills/scripts surface, with the fable mandate noted above).

Source

Observed during the 2026-08-11 engine-core seat takeover (session session_01PemxJmMZT9JphrsNx886Ce); maintainer prompt quoted above. Dedup: searched pm:seat size/compaction cards — none exist.