[EHotwagner]

fdev: the finite settled usage and review census for #4823 (a child of #4639). It covers the window [2026-10-19T18:19:36Z, 2026-10-23T00:00Z), as fdev's rescope sets it.

What this PR adds

• docs/reports/2026-10-23-usage-census-4823.md: the report, with its method, coverage, Gaps and Unknowns, and one row per member.
• docs/reports/evidence/2026-10-23-usage-census-4823.json: the compact per-member result as column/row tables, under the 64 KiB telemetry-evidence cap. It holds no session identities and no transcript content. The private-repository members are redacted, including their numbers, SHAs and dates, and are listed last.
• The Atlas record github/settled-usage-review-census-method (lesson, digest 2d46924a…), mirrored into both seeds. The driver manifest is regenerated.

Headline results

• Population: 134 merged PRs and 34 closed Project 3 items. Both member lists are frozen by digest.
• Review: 44 PRs have an independent verdict bound to the exact final head. The other 90 are unsampled, each with a stated reason.
• Required checks: 131 pass, 1 fails (Rendering#1464, an admin bypass recorded in that PR), 1 repository has none configured, and 1 is unknown (the private repository).
• Actions job wall time: 283.6 h in the PR phase and 97.6 h on merge-commit pushes, from the Jobs API.
• Native tokens on the coordinator host:
  • Claude Code: 1,777,676,869 tokens.
  • Codex: 5,304,222,463 tokens.
  • 65,125,639 tokens are attributed by branch join to 3 PRs. Everything else is unattributed.
• Unknown: billed cost, other hosts (including Home), and pruned transcripts.

Effect boundary

• The usage reader is the source build of .github at 61526e94, which fdev admitted for this census only. It ran read-only, outside any repository.
• The receipt store was a scratch directory.
• No UTEL store, telemetry database or receiver was opened or written.
• The Jobs API pass was paced to leave headroom in the shared REST quota.

Checks run locally

• python3 tests/atlas/check.py --baseline <merge-base>
• scripts/skill-view-parity.sh (AGREE)
• scripts/generate-driver-manifest --check
• tests/skill-quality/run.sh
• python3 scripts/check-telemetry-git-safety.py --base <merge-base>
• git diff --check

Issue 4823 stays open after this merge: native acceptance and closure stay with the root, and the parent outcome (4639) also stays open.

Refs #4823

🤖 Generated with Claude Code

[EHotwagner]

fdev 2026-10-23: Independent review record. fsgg-critic-best: ACCEPT on exact head ec6f56548264e93c1bda2d4ce4563b7c6570bcce.

The candidate was reviewed against the #4823 Done-when and fdev's rescope.

• Round 1, scratch draft: REQUEST CHANGES.
  • A private-repository member leaked into the public report.
  • Review false negatives: hand adjudication now covers every PR whose record mentions a review.
  • Stacked PRs double-counted their PR-phase runs. Each run now goes to exactly one member, by head branch and then by which PR was open.
  • Non-push runs on merge SHAs were counted. Only push runs count now.
  • Failed Jobs API reads were never retried.
  • The attribution join parser was not disclosed.
  • The digest description did not match how the digest was computed.
  • One route marker was misreported.
• Round 2, df423ae6: REQUEST CHANGES. All of the round 1 findings were fixed. The close times of the two redacted private Project 3 items, together with the public member digest, still let their numbers be recovered by brute force.
• Round 3, ec6f5654: ACCEPT.
  • The redaction is complete, and redacted members are listed last.
  • The values are unchanged from round 2.
  • Atlas record 55adfd09… re-cites the report and evidence at bf029bc6.
  • The force-push gap is stated with counts.
• Nonmaterial (not addressed): the evidence JSON abbreviates the route marker as rd/v1 with no legend. The report table shows the full form.

Required checks on ec6f5654: all pass. The PR has 54 passing checks and 2 skipped. The cancelled declared-closes run was superseded by a passing rerun.

[EHotwagner]

Coordination takeover: preserve the active census writer

The user instructed this Codex coordinator to take over, stating the previous coordinating session is dead. During same-PR recovery, a surviving Claude-owned process changed this PR's existing worktree from the reviewed ec6f5654 to local 529198a4c108260d0959ecd9d71955fb49527b26 and continued updating Atlas catalogs/manifest and a new 2d46924a… record. Native PR head was still ec6f5654 at readback. The new root worker stopped without editing, committing or pushing; all surviving work is preserved.

Existing census writer: finish or safely stop your currently owned bounded preparation; do not start a competing PR or recapture census data. Reply on this PR by 15:20 UTC with current exact head/tree, dirty paths, tests, original ownership/operation and whether any push is in flight. If more time is needed, provide the next checkpoint. Root will independently reconcile the result and continue this same PR.

Root: waiting on that source-custody result for the affected worktree, while Home receiver and other independent work continue. No process is killed, no files are overwritten, and no old reservation is treated as released. Existing review remains valid only for its exact ec6f5654 scope; final-candidate review and required CI still apply.

[EHotwagner]

fdev 2026-10-23: Independent review record, round 4. fsgg-critic-best: ACCEPT on exact head 684f44909377021aa63118fec46b141f71bf1fba.

Main had moved: #4833, #4844 and #4846 added Atlas records, which conflicted with this PR's catalogs and manifest. The branch was rebuilt on fcbb8eac as two commits:

• Report and evidence. These are byte-identical to the content accepted at ec6f5654.
• Atlas record 2d46924a…. Against 55adfd09 it differs only in recordedAt and in the revision its report citations point to, now 529198a4. The cited file hashes are unchanged.
  • Each of the 3 catalogs is main's catalog plus this record: 383 records, with nothing removed.
  • The manifest changes only the tree hash, the catalog hash and the new record entry.

All required checks pass on 684f4490: 54 passing and 2 skipped. That includes routine-eligibility and declared-closes. The PR marker is bound to this head.