# INFRA-REGISTER sweep — SEMAPHORE #242 close (doyle, DRAFT 2026-08-29, pre-golden)

Full-register pass done (72 entries; delegated enumeration, load-bearing claims
re-measured in-tree). Executes at milestone close AFTER golden green + release cut —
nothing lands on main before the tag (tag==main==tested parity).

## Measured corrections to carried state

- **IR-57 figure correction (310/308→302/300) is ALREADY LANDED inline** in the entry
  ("Corrected todlando 2026-08-29, measured both ways in-tree", register ~line 2964) —
  the "rides next register batch" note in SEMAPHORE-242-GRILL riders + my commune is
  STALE. No batch item remains. Remaining IR-57 close-work: append today's #241
  assembly first-use line (5 MATCH / 1 LOSS reconciled — ASM-241-GATE-VERDICT.md) as
  evidence; entry stays OPEN as living procedure.
- **IR-37 flag from sweep agent is OVERCAUTIOUS:** kitsubito's missing treqs binary
  (ASM_treqs 127) is BOX tooling; IR-37's consume lane is spt-core CI (GitHub runners),
  a different surface. Golden's treqs leg runs on runners and was green all week. Keep
  the kitsubito install-or-drop-leg item as its own small rider, do NOT reopen IR-37.

## Close-out composition (execute at close)

1. **IR-69 follow-through — EXECUTED 2026-08-29 ~20:30Z, FINDING:** close-out re-census
   at `ec6da9b0` (independent meter, containment-classified) found the generator's
   truncation blind spot LIVE — cli.rs scanned only to line 3007 (first module-opening
   `#[cfg(test)]`); **338 TOKEN-shaped shipping sites after it never censused, still
   bare**; + SELECTION_PROBE space-family residual (membership unverified) + post-cut
   minted sites in delivery.rs/main.rs. Full write-up
   EMISSION-RESIDUAL-CENSUS-FINDING.md; board request filed via alchemy (residual
   conversion + generator fix + enforcement-at-seam); todlando/hertz reconciliation
   requested against sealed artifacts. **IR-69 stays OPEN with addendum** — does not
   close at this sweep. Common-word matcher ruling still owed (unchanged). The shipped
   #241 conversions (censused 486, 495 removals) are unaffected; golden green stands.
2. **Daemon-leak cluster → hertz's queued fixup lane carries it:** IR-7, IR-17
   (activity_link_push_e2e named directly), IR-20, IR-34, IR-35, IR-63. His lane brief
   cites these entry ids so each leaves the register lane-linked.
3. **Disk-floor cluster:** IR-59 already slated next-intake tooling rider (with IR-46
   golden-preflight face). NEW FILING at sweep: kitsubito 21,643 /tmp endpoint homes
   (hertz's measurement) — sibling of IR-31/IR-59/IR-64 family, box-side accumulation.
3b. **Disk-floor family write-up additions (from #242 golden attempt 1, deployah):**
   (i) per-job floor figures are SEPARATE instants — quote per job (test 26521477120,
   n1-gate 26535051264, 8s apart, one run); (ii) CORRECTED TRAP (deployah measured,
   retracting his earlier "attempt-2 hides jobs" version): the default `/runs/<id>/jobs`
   endpoint stamps `run_attempt=<current>` on EVERY job including ones that never
   re-ran (4 of 8 mis-assigned on #242 golden) — it is the RUN's attempt, not the
   JOB's; a stale green can read as re-earned. `started_at` is the discriminator,
   `/runs/<id>/attempts/N/jobs` the authority. Also: `rerun --failed` re-runs DEPENDENT
   jobs (twohost re-ran despite green attempt-1); (iii) rerun preflight reads RAW free
   bytes with the CI step's own predicate, not display GB; (iv) rerun greens must be
   checked NON-VACUOUS at step level (checkout SUCCESS + workload steps ran) before the
   floor RCA counts as closed by evidence.
3c. **NEW FILING at sweep — literal-first floor sites OFF the golden path (deployah
   flag, 2026-08-29, widen lane):** ci.yml (5 tags) and release.yml (3 tags) still
   read the free-space floor PRE-checkout — the exact REQ-CI-FREE-SPACE-PREFLIGHT
   trap whose "other jobs keep the literal-first shape" clause cost #242's cut two
   golden attempts (n1-gate). golden.yml is fully reshaped (ac7d2609, both jobs,
   checkout-protective reaps verified to stay pre-checkout); these two workflows were
   deliberately NOT widened off the back of a golden.yml ruling. File as rider on the
   disk-floor cluster: review each site's comment for checkout-protective purpose
   (the ac7d2609 method), then reorder or record why not. RIDER 2 (r4 LNK1318,
   2026-08-30): the floor is asserted once at job START; the Windows test job's
   docs-drift build is the LAST heavy step, ~28 min in, running at the job's
   internal LOW-WATER (r4: floor read 65.1 GiB honest at 23:07, box at ~37-39 GB
   by the 23:35 link failure; LNK1318 on PDB write = documented low-disk symptom;
   incremental-corruption candidate retired by preservation snapshot — no xtask
   state existed at failure). Durable fix = floor RE-READ (or reclaim) immediately
   before the docs-drift build — a WITHIN-job sibling of the across-job fix, rides
   the next workflow commit. Plus incremental hygiene (dir was 4.56 GB) as
   disk-pressure maintenance, and the box's ~51 GB non-recovering post-run
   consumption is the IR-59 pool-weight face — classify at the audit slot.
3d. **From the r2/r3 RCA (RCA-242-R2-LINUX.md, rides respin/post-cut — link when
   landed):** (i) BRAIN_PHASE breadcrumbs in the silent BRAIN_UP→ready window
   (todlando, pulled ahead of cut); (ii) hertz 4-cell fixup: false precondition
   message + progress-bounded (not bigger-clock, #235 shape) readiness wait —
   NOTE IR-17 names activity_link_push_e2e directly, check overlap before
   lane-linking — AND (deployah measured 2026-08-29): IR-17's 2026-08-03
   "EMPTY brain stderr" specimen is a METER ARTIFACT, not a silent brain — the
   panel capture landed 2026-08-19 (ab657626 + e6e188e1), 16 days after the
   specimen; current bits label absence explicitly, so empty-vs-present cannot
   recur as a distinction. IR-17's entry text owes this correction at sweep
   (correct by replacement, keep the mechanism); (iii) BOARD REQUEST to mint at #243 filing batch: move
   current_exe_hash() off the ready path (behaviour change to ADR-0018 Q7 ratified
   readiness semantic — operator triages; cited to RCA).
4. **Kitsubito treqs rider (new, small):** install traceable-reqs on kitsubito OR drop
   the leg from Linux gate drivers; a 127 leg is exactly the wrapper-green shape IR-60
   documents (deployah co-flagged at intake).
5. **Pool/worktree idle-window audit:** IR-14 (.worktrees population — ~85 dirs
   currently), IR-26 (announce half), IR-27, IR-49 — between-milestone window opens at
   release. Compose as one audit slot.
6. **Small riders:** IR-11 (headless column, trivial), IR-33(a) (clap stack smoke).

## Stale-note hygiene at close

- Update grill riders line (IR-57 "queued" → landed inline).
- Commune correction: v0650-arc / next commune must NOT carry "IR-57 correction rides
  next register batch" forward.