Witness cost class: import-closure pollution — zero confirmed exemplars

[gunbai-bot]

Summary

Work item: node://adhoc-0e6f8ea5-158 ("Witness cost class: import-closure pollution — root-cause 2-3 exemplars, land a reusable closure-narrowing repair, restore the modules with an exact planned/problem/collateral delta"), child of royal-hawk-392.

Honest result: zero confirmed exemplars. Three candidates were investigated; all three were falsified as import-closure pollution, each remeasured against the required floor's own [floor-witness-slow] log line (not a standalone harness run, which was found to overstate cost ~2-3x due to a once-per-claim vs once-per-MultiEntryIndex amortization difference).

The three candidates

1. extdeps_scope_placement_gate_loudness_witness (PR Take the extdeps-scope loudness probe out from behind a 718-module import #8432's 718→62 module import-closure narrowing). The narrowing is real and DESIGN.md §3-legitimate, and standalone resolve time dropped (30.7s → 1.0s). But the floor's own last post-repair line (red_seed_runner_failure_detail_projects_located_receipt, run 32177951514, head 59653f569e, a confirmed descendant of the repair commit ff56305b731) shows 75926ms, BUDGET-REFUSED against the 1552ms ceiling, +1.00GB RSS growth (the worst single claim in that run) — statistically unchanged from its pre-repair floor history (42983-97260ms). The module is now excluded from the floor entirely by #8457's test.claim.long. rename, so no current floor line exists at all. Right repair, wrong quantity measured.

2. Class B rust_test_fixtures rows 1-2 (declared_import_closure_lists_rust_import, ambient_pool_provider_preserves_listed_import). A paired same-host, same-dispatch comparison found the narrow probe-backed entry and the heavy transport-backed entry statistically indistinguishable (~70s/~80s standalone, both). The floor's own line (24674ms/31603ms) confirms the same non-divergence at its own lower, amortized magnitude.

3. dissolution_census unbound-literal witnesses. Root-caused to a different mechanism: compile_dag_rust_emit_check's uncached path rebuilds a full corpus-rooted module index per call, memoized only under a real --required-floor run — a first-toucher shared-fill signature (42129ms vs 825ms, same module, no import-closure difference). This is class-2 per-claim memoization, not closure-narrowing scope, and out of this lane.

The common object

All three point at the same underlying cost: a fixed, corpus-wide load/index term paid once per MultiEntryIndex rather than once per entry's import closure. This corroborates — does not re-derive — DESIGN.md's floor-shared-compute-memoization open thread on build_both_closure_edge_index, whose closure-size-independence this lane's measurements are consistent with on three independent subjects.

What's retained (not cost repairs)

• src/v2/workflow/class_b_import_closure_probe.dag (new): rows 1-2's shape+call, relocated out of class_b_import_closure_transport.dag per the same DESIGN.md §3 relocation-not-refork discipline already used for src/v2/compiler/inferred_tree.dag (also included here, previously split from 04_infer.dag). Both splits genuinely narrow dependency footprints and are legitimate modeling corrections — neither is claimed to buy measured cost.
• src/v1/stage0/src/cli_run.rs: adds CLASS_B_PROBE_REL to CLASS_B_GATE_INPUT_ENTRIES — a real collateral correctness fix so the gate's input closure covers the relocated module.
• Three .dag note strings updated with the corrected, floor-sourced magnitudes and explicit UNCONFIRMED/FALSIFIED framing.

The reusable deliverable

The paired same-process, same-host narrow-vs-heavy comparison methodology, baselined against the floor's own CI job log rather than a standalone run:

gh api repos/:owner/:repo/actions/runs/<run_id>/jobs --jq '.jobs[0].id'
gh api --allow-escape-sequences repos/:owner/:repo/actions/jobs/<job_id>/logs

This is the two-sided acceptance bar for any future import-closure-pollution candidate in this population: cost must collapse when the closure narrows AND not collapse under repetition in one process — measured against the floor's own line, never a standalone run.

Test plan

• cargo fmt --all --check (pre-commit/pre-push hooks passed)
• Remote release build of v1-compiler compiles cleanly with these changes applied (verified via background dispatch; the one failure was a CLI-flag usage error in my own verification invocation, not a compile failure — the Finished release profile line preceded it)
• Full floor pass — not run from this session per the resource-sharing constraint below; the module changes are note-only or dependency-closure-only, no test-body semantics changed

Note: a full required-floor run was attempted from this session and OOM-killed (session containers share sessions.slice's aggregate memory cap; a single claim on this population alone grows RSS to 10+GB on the CI runner). Per-identity floor lines were instead pulled from existing CI job logs via the gh api route above, which is the now-established correct method for this class of measurement.

🤖 Generated with Claude Code