Run the resolution-divergence census in-process

[gunbai-bot]

Summary

Run the resolution-divergence silent-pick census in the claim executor process over the parent-owned shared MultiEntryIndex. The CI witness no longer launches a subprocess, and no child performs a second closure/whole-tree resolve. The standalone resolution_divergence_census CLI remains a thin cold adapter over the same projection and verdict core.

Before changing that mechanism, add fsynced, kill-safe phase receipts for parent plan resolution, child launch, child whole-tree resolution, parent shared-index closure resolution, census traversal, and output projection. Receipt rows also carry host load and cgroup pressure and stream to CI logs, so a killed row still locates consumed wall. The existing 900-second row budget and witness selection are unchanged.

Fleet acceptance receipt

Falsifier run 30735812900 on arm64 runner srv1-04-1785650374-2886294 executed resolution_divergence_silent_pick_gate_keystone_holds uncensored and passed at 244,406.6 ms (row-cost total 245,380 ms), 34% of the 720-second margin bar and 27% of the unchanged 900-second wall.

Phase	Receipt
parent plan resolve	17,072 ms
child launch	Skipped
child whole-tree resolve	Skipped
parent shared-index closure resolve	240,063 ms; 1,764 modules
census traversal	3,434 ms; 62,163 sites
output projection	733 ms; clean gate verdict

This proves one census closure resolve in the parent and none in a child. The frozen-tree control independently proves byte-identical stdout/stderr/exit projection and the same gate verdict before and after the cut.

Conditions: ten other CI runs were active at dispatch and no other falsifier was executing. During the target phases, host load moved 8.73 to 10.35, CPU PSI some avg10 moved 0.09 to 0.00, and memory PSI stayed 0.00. The job-scoped cgroup peak was span-calibrated from 8.26 GB before the floor to 16.11 GB after it because that runner did not permit resetting memory.peak.

The sample SHA (2070e3ddaded) predated the merged REST ContentHash repair #7665, so the full workflow was red on those six unrelated rows even though the target passed. Implementation head d669d5f0359 merges main including #7665 and is fully green in PR CI run 30746485038: build, regen, artifact healing, and floor CI all pass; merge state is CLEAN.

GitHub Actions concurrency permits one executing and one pending falsifier; a new cron evicts the pending entry even with cancel-in-progress: false. The one authorized current-head retry, 30746679393, received zero jobs and was evicted at 13:35:10Z by scheduled run 30750244803 created one second earlier. Per the acceptance-owner sequencing decision, no further manual retry competes with the production cold control: the 244.4-second valid sample is the pre-merge receipt, and the row-cost/phase receipt remains enrolled in every scheduled falsifier after merge as the continuous production check.

Remaining measured cost

The duplicate resolve is removed, not made warm: both child phases are skipped, while the one remaining parent resolve accounts for 240.1 seconds and traversal for 3.4 seconds. Immediately before the target, floor-drain retention had reduced the shared typed cache to 8 modules; two preceding entries raised it only to 47 before resolving the 1,764-module closure. The target's retention receipt carries schedule_evictions=557, graph_evictions=321, and retention_unknown=0.

That is the measured firing condition for DESIGN's existing floor shared-computation memoization thread: “Retention PARKED unless M2-aware floor shows pressure.” This PR records the receipt but does not widen into floor-wide retention work.

Source identity wall

The first fleet execution exposed absolute-vs-repo-relative SourceFile identity: identical bytes were loaded into two distinct Rc objects. The in-process path now seeds its closure from the parent index's owned source objects, and the frozen control asserts Rc::ptr_eq for every source. Content equality is deliberately not the oracle because it passes for the bad pair.

This is a site-local wall. The class-wide next rung—canonical identity at construction, so path spelling cannot create two objects—is tracked by node://adhoc-0c214548-ded; the pointer-identity control stays enrolled when that higher rung lands.

Hand-Rust scaffold receipt

Final review head 71eac33bb98 declares the entire in-process census/phase-receipt surface as DESIGN §7 seed-retained work at the witness boundary. The measured final Rust census is 765 added and 30 removed lines across five seed files, net +735; the declaration explicitly records growth rather than claiming a shrink. One Scaffold { dissolves_to: SingleAuthority } binds the deferral to the v1-exit lane and concrete ROADMAP row v1-zero-hand-maintained-rust, and the keystone executes assertions over the Scaffold shape, census receipt, singular trigger, and row identity.

Post-review hardening also brackets every fallible shared-index setup step and unconditionally disables silent-pick observation before propagating an error, so a closure/receipt refusal cannot leave global telemetry armed for later work.

Verification

• cargo test -p v1-compiler --lib shared_index_cut_preserves_frozen_census_output_and_verdict -- --nocapture — passed on final Rust head dc92a0d (final review head changes only checked scaffold census text) (BuildBuddy)
• cargo test -p v1-compiler --lib phase_receipt_persists_started_before_completed_with_distinct_timing — passed
• cargo test -p v1-compiler --lib parent_and_shared_index_telemetry_join_without_cache_hit_duplicates — passed
• cargo check -p v1-compiler --bins — passed on final Rust head dc92a0d (final review head changes only checked scaffold census text) (BuildBuddy)
• cargo check -p v1-compiler --lib — passed with telemetry refusal cleanup (BuildBuddy)
• cargo run -p v1-compiler --bin regen_stage0 -- --verify — regen_divergence_count=0
• cargo fmt --all --check and git diff --check — passed
• Final-head GitHub checks — build, regen, heal-generated-artifacts, and floor CI passed on 71eac33bb98 (run 30751327779)

Worker attestation

• Title describes the change.
• Body summarizes what and why.
• Tests, fleet timing, contention conditions, and outcomes are recorded above.
• This dashboard work item is not a GitHub issue, so no Closes #N directive applies.
• All commits on this branch were created in this session.
• No secrets, credentials, or large binaries are included.

[gunbai-bot]

Addressed review 47362 in 186003f. The in-process census/receipt surface is now a declared DESIGN §7 seed-retained Scaffold row at the witness boundary. Its receipt records the measured Rust census (758 added, 30 removed, net +728—explicitly counted growth, not a shrink), enumerates the added seed surface, and binds one checkable dissolution trigger to the v1-exit lane’s concrete v1-zero-hand-maintained-rust ROADMAP row. The falsifier keystone now executes assertions over the Scaffold shape, census receipt, singular-trigger wording, and roadmap-row identity. regen_stage0 --verify on 186003f reports regen_divergence_count=0; the final-head CI build and artifact-heal jobs are green, with regen/floor still running.

[gunbai-bot]

Correction to my review 47362 response: the subsequent telemetry error-path cleanup changed the final Rust census by seven net lines. The final checked receipt at 71eac33 now records 765 additions, 30 removals, net +735. The Scaffold trigger and ROADMAP binding are unchanged.