Re-derive the effect floor inventories against current lane lowerings

[tamashi095]

368 — Re-derive the effect floor inventories against current lane lowerings

Split out of the #349 efficiency-audit tracker, Wave 0 (correctness and gate integrity). Rows: LANE-13.
This issue is self-contained: implement from this body alone. Do not re-audit #349.

Assignment

Implementer	Codex
Verifier	Fable 5.1 (reads the diff, reruns every gate below, posts the tracker note on #349)
Branch	codex/368-lane-13-floor-recount off main
Land after	LANE-4 (#367) — the finding says one recount must cover both
PR	one PR, squash-merged, title prefixed with the row id(s)

Operating procedure (complete; no other instructions exist)

Repo: /Users/bl/Documents/GitHub/misofm/engine (GitHub misofm/engine). N is this issue's number.

0. Claim. gh issue view N --json labels must not contain claimed; if it does, stop, another agent owns it. Every issue in the "Land after" row must be CLOSED (gh issue view <dep> --json state); if not, stop and say which one is open. Then gh issue edit N --add-label claimed.

1. Worktree. Never work in the main checkout.

git -C /Users/bl/Documents/GitHub/misofm/engine fetch origin
git -C /Users/bl/Documents/GitHub/misofm/engine worktree add ../engine-N -b <branch from the Assignment table> origin/main
cd /Users/bl/Documents/GitHub/misofm/engine-N && export CARGO_TARGET_DIR=$PWD/target

Read AGENTS.md and docs/REALTIME_DEPENDENCY_POLICY.md once before editing.

2. Scope. Implement only the finding(s) in this issue, following any "Reconciled scope" / "Notes for the implementer" section first and the finding's "Proposed change" second. Locate code by the named functions; line numbers are from main 4797a544. Anything else you notice goes in the PR under "Seen, not done" and stays untouched. No reformatting, renames, dependency bumps or new tooling. Where the finding lists ordered options, try them in that order; an option marked class B or owner ruling is never implemented: if you reach it, post the evidence on this issue and stop.

3. Null result. If the premise no longer holds on current main (code already matches the proposal, or the named code is gone), do not invent a change. Comment on this issue with a one-paragraph null note, remove the claimed label, and stop.

4. Evidence, all required.

• The exact "Verification gate" commands from the finding, output saved.
• cargo test --workspace in the worktree and on origin/main; pass counts must match unless this issue adds tests (state the delta).
• Any script you edited: also run its scripts/test-<name>.sh mutation harness.
• Render-path rows only: scripts/check-realtime-policy.sh, the bit-identity fixture suite named in the finding, and one run of scripts/run-console-benchmark.sh on the named row. Report before/after once; never tune.
• Codegen claims (an instruction that should disappear): attach before/after llvm-objdump -d (or cargo asm) for the named function.

5. Deliver.

• Commit: subject <row-id>: <what changed>, body listing the commands run, last line Refs #N.
• gh pr create --base main with sections: Summary; Rows closed; Gate output (fenced); Before/after; Seen, not done; Skipped (any step not done and why). Add Closes #N only when every row in this issue is done.
• Do not post on Engine efficiency audit: render-path cost, DSP/SIMD peak-efficiency, DRY, and crate-boundary findings #349; the verifier does that. Do not git push --force, do not touch main, do not edit outside the worktree.
• Reply with the PR URL and the Skipped list.

Finding (verbatim from #349)

LANE-13: floor-accounting gap term (a) is already closed in code but the ruling's tables still price it as open

• Category: architecture
• Class: N/A (accounting)
• Severity: low
• Location: crates/lane/src/wide_impl.rs:39-64 and :263-271;
  docs/rulings/effect-floor-accounting.md:128-129,134-136,183-186; tools/bench/src/floor.rs;
  scripts/console-benchmark-record-lib.jq
• Current code:

// crates/lane/src/wide_impl.rs:264
fn select(m: Self::Mask, a: Self, b: Self) -> Self {
    m.select(a, b)   // wide 1.6.1 -> blend_varying_m256 == vblendvps on avx
}

• Problem: the brief's gap term (a) — "bitselect lowering to 3 bitwise ops instead of vblendvps" —
  no longer exists. Verified in wide-1.6.1/src/f32x8_.rs:148-158: select is
  blend_varying_m256(if_false, if_true, mask) under avx, i.e. one vblendvps, and
  f32x4_.rs:236-248 is blend_varying_m128 / v128_bitselect / vbslq_f32. The semantics argument is
  sound (blendv reads only the sign bit; every mask the trait can build is all-ones or all-zeros, and
  the four mask constructors in kernels::builtins are themselves ordered compares). But
  effect-floor-accounting.md:134-136 still says "On the emitted code it is three — LLVM builds
  (a & m) | (b & andnot m)" and the compressor gap table at :183-186 still attributes +24
  instructions per channel-frame to it. An implementer reading the ruling will re-do work that is done.
  Separately, wide_impl.rs:108-117 records that max/min are now one instruction on x86 and wasm
  while the floors still price them at two — an owed recount that has not happened.
• Proposed change: re-derive the compressor/EQ/limiter/builtins inventories in
  effect-floor-accounting.md against the current lowerings (select = 1 emitted, max/min = 1 on
  x86/wasm and 2 on NEON, exp2_int = 2 after LANE-4), and re-issue the floor constants in
  tools/bench/src/floor.rs and scripts/console-benchmark-record-lib.jq in the same change, as that
  ruling requires. Do it after LANE-4, so one recount covers both.
• Expected effect: the standing authority every sealed console record is read against stops being
  conservative by an unquantified margin; gap term (a) is struck from the open list.
• Rulings / constraints checked: effect-floor-accounting.md:108-117 explicitly says this recount is
  owed and must be an owner-visible change; this finding is the trigger, not an unauthorised re-pricing.
• Verification gate: bash scripts/run-console-benchmark.sh against the re-issued floors; the
  record validator (scripts/console-benchmark-record-lib.jq) must accept the sealed records unchanged
  apart from the re-priced floor fields.

Constraints (from #349, binding)

• Unfused (a*b)+c everywhere; bit identity across scalar/W4/W8/wasm. A class A row moves no rendered bit. If the row turns out to need one, stop, re-class it B on Engine efficiency audit: render-path cost, DSP/SIMD peak-efficiency, DRY, and crate-boundary findings #349, and do not implement.
• No new crate names, no renames for their own sake, no runtime SIMD dispatch, no unsafe outside the existing allowlist without a ruling.
• Render path is zero alloc/free/lock/syscall/I/O/log/panic (scripts/check-realtime-policy.sh, docs/REALTIME_DEPENDENCY_POLICY.md).
• One measurement per row where the finding names one, recorded, not chased. No tuning loops.
• Line numbers above are from main 4797a544 and may have drifted; locate by the named functions. If the premise no longer holds on current main, close this issue with a one-paragraph null note and append the same note to Engine efficiency audit: render-path cost, DSP/SIMD peak-efficiency, DRY, and crate-boundary findings #349.

Done when

1. The finding's Verification gate above is green, with the command output attached to the PR.
2. cargo test --workspace passes with the same pass count as main, plus the named policy scripts.
3. A one-paragraph note on Engine efficiency audit: render-path cost, DSP/SIMD peak-efficiency, DRY, and crate-boundary findings #349: row id, commit hash, what changed (or the null result).

Spec mirror: .github/ISSUE_SPECS/368-lane-13-floor-recount.md.

---

Reconciled execution brief (2026-09-04; supersedes conflicting assignment/procedure text above)

Astra brief: #368 / LANE-13 — current-lowering floor accounting

Decision: APPROVED scope, root mirrors this brief into the local #368 spec and synchronizes the GitHub issue before implementation. Reviewed main 8792698. #367 is CLOSED; #378's owner ruling explicitly defers native AArch64 and unblocks #368. #368 is claimed by root for this workflow. This is the next eligible Wave 0 slice; no dependency exception is needed. #375/#389/#390 own unrelated live changes and must not be touched.

Smallest closable outcome

One accounting-only PR makes the standing four kernel inventories and the Rust/jq floor composition agree with current lane operations, removes already-closed select lowering from the live gap list, proves the repriced records change only floor-derived data, and captures one honest descriptive console invocation. No DSP implementation or optimization is part of this issue.

Frozen scope and authority

Read #368's existing body, AGENTS.md, docs/REALTIME_DEPENDENCY_POLICY.md, docs/rulings/effect-floor-accounting.md, docs/rulings/unfused-multiply-add-audit.md, crates/lane/src/wide_impl.rs, relevant compressor/EQ/limiter/builtin kernels, and tools/bench/src/floor.rs. The user's current assignment overrides the old model table: Astra briefs and reviews, Luna gets exactly one implementation attempt, Sol retries only if that attempt fails. Root owns Git checkpoints, pushes, PR creation, merge and issue synchronization. Implement in a new isolated worktree off synchronized origin/main; never use another active worktree or edit the shared main checkout.

Allowed files: docs/rulings/effect-floor-accounting.md; tools/bench/src/floor.rs; scripts/console-benchmark-record-lib.jq; focused additions/adjustments in scripts/test-console-benchmark.sh and existing floor.rs tests; a comment-only resolution of the owed recount in crates/lane/src/wide_impl.rs if needed; the #368 spec and brief/evidence record; new artifacts/issue368-floor-recount evidence. A minimal --issue368-floor-recount arm and usage text in scripts/run-console-benchmark.sh is authorized because all existing namespaces are historical one-shot authorities. Do not refactor that runner. Any other file needs a brief amendment first.

Explicit exclusions: rendered arithmetic, operation order, SIMD implementations, effect parameters, buffers, graph/compiler/protocol/host dependencies; class B changes; throughput-probe reruns; new benchmark framework; floor machine calibration; mono-collapse pricing decision; re-pricing the intentionally rounded reduction amortization; generic Rust/jq key-set tooling (TOOL-8); native AArch64 qualification. Keep BANK_WIDTH=8 and OPS_PER_CYCLE=3.7; state their historical host origin and do not imply they were calibrated on the current machine.

Derivation contract

1. Recount expressions, not prose constants. For the current x86/Wasm lowerings max/min cost 1; select already had a 1-op floor, so closing the emitted select gap does NOT subtract more floor ops. exp2_int_in_range is the post-Add exp2_int_in_range so fast_db and lane_math skip the dead clamp #367 two-operation synthesis on the admitted domain; callers retain their clamp. Do not label public clamped exp2_int a two-op method. NEON is unsupported/deferred; describe its source-level intended compare/select shape only with that qualification, never as verified emitted behavior.
2. Re-derive each line for compressor, EQ, limiter, builtins; identify exact called function for each changed line. Fractional per-channel counts are valid: a single shared stereo max saves 0.5 op per lane-sample. Do not round shared work up to an integer. Preliminary delta-only cross-checks, NOT acceptance oracles: compressor 94 - 12.5 = 81.5, limiter 138 - 8.5 = 129.5, EQ 51 and builtins 69 unchanged. Independently trace the source; report any discrepancy before broadening scope. Whole strip and derived row compositions must follow the justified totals.
3. Distinguish current authority from dated evidence. Keep historical measured cycles, timing, captures, and sealed artifacts untouched. The original +24 emitted-select gap and old disassembly counts must be explicitly historical, not asserted of current main. Add a clear current recount/closed-gap statement wherever a reader could mistake a historical optimization suggestion for open work. Recalculate current derived tables; historical tables may retain original pricing when clearly dated and labeled. Resolve the stale max/min recount debt, while preserving the separate unresolved mono-collapse ruling.
4. Do not claim instructions disappeared in this PR: it changes accounting only. Cite current source/lowering and existing Add exp2_int_in_range so fast_db and lane_math skip the dead clamp #367 evidence for prior codegen changes. If making a fresh emitted-code claim, capture and identify the exact current function/artifact disassembly; never re-label old captures as current or transplant their measured instruction totals into the new table.
5. Rust and jq must independently express the same justified inventory and compositions. No widened tolerances, guessed clocks, renamed schema, changed output digests or weakened validators.

Evidence gates (before PR)

• Root records baseline cargo test --workspace on origin/main and implementation worktree (PATH=/home/bl/.cargo/bin:$PATH). Candidate must pass with identical baseline count plus any explained focused tests. Save summaries/logs. Baseline infrastructure failures are candid blockers, not silently waived gates.
• cargo test -p bench for inventory/composition tests; bash scripts/test-console-benchmark.sh for validator acceptance and red mutations; bash scripts/check-realtime-policy.sh; formatting and proportional required policy gates. If the runner changes, run its existing console mutation suite and operator preflight, not an invented timed rehearsal.
• Add a focused counterexample to catch stale old floors in Rust/jq tests. Include a fractional stereo-link assertion and at least representative compressor+limiter compositions so a mistaken integer recount cannot pass. These should discriminate the accounting claim, not pin documentation prose.
• Historical-record comparison: copy representative sealed console records containing real cycle/floor columns to a new temporary/evidence namespace (never rewrite their originals). Update ONLY floor_cycles_per_lane_sample, percent_of_floor, and isolated_percent_of_floor according to current compositions and existing measured clock/timing/control rows; any other changed field must be explained and authorized. Compare JSON after removing precisely those fields for exact equality. Run the actual record and aggregate validators on the repriced set. Include standalone/ragged/unpriced rows when present. Prove stale old floor fields fail for affected rows and unrelated malformed fields still fail. Existing no-cycle records continue validating unchanged. This establishes the finding's 'sealed records unchanged apart from re-priced floor fields' without falsifying provenance.
• Freeze validator/workload and finish all non-timed gates FIRST. Preflight arguments, unique output namespace, output persistence, overwrite refusal, exit semantics and available counters without starting the timed workload. Then root authorizes the single benchmark invocation: bash scripts/run-console-benchmark.sh --issue368-floor-recount, exactly one warmup and two measured rounds under the existing runner contract. Use MISO_ENGINE_BENCH_ALLOW_UNCONTROLLED=1 only when host controls are unavailable and disclose the exact reason. Do not invent a core clock if perf is unavailable; omitted cycle fields are the supported representation, and the repriced historical-record proof above still exercises actual floor fields.
• If post-workload tooling fails: preserve raw/stdout/stderr/disposition, do not rerun or tune, report the failure immediately. One bounded tooling repair may be separately scoped; the issue cannot claim the original benchmark gate green while it is red. Root must rescope/split the qualification remainder rather than allowing this accounting feature to become a runner project.
• This is not an effect optimization loop: do not claim runtime speedups from floor repricing or describe historical isolated cycles as a current measurement. Where reporting residual gaps, give actual record-derived values, the new derived floor and a named reason; use 'not investigated' if necessary.

Attempt/checkpoint/delivery rules

Luna executes one coherent implementation pass and focused verification, then pauses and sends exact paths plus evidence to root for an immediate compiling checkpoint. No second implementation tranche while that checkpoint is waiting. Root commits/pushes according to active delivery mode. Astra reviews the resulting PR against this brief and named gates before merge. A failed Luna attempt goes to Sol for a bounded revision (at most two Sol implementation attempts total); each gets one adversarial Astra verdict. After three failed attempts, stop and rescope; no disguised retry. Do not weaken any gate.

PR title starts LANE-13 and describes accounting. Include Summary, Rows closed, Gate output, Before/after (old/new inventories; no speed claim), Seen/not done, Skipped. Root closes #368 only after PASS and its evidence commit is upstream; then verifies CLOSED and posts #349's row/commit/outcome note. #349 remains an audit tracker, not authority for a cross-cutting cleanup branch.

Attempt 2 decision/evidence record (2026-09-04)

Astra's attempt-1 verdict was FAIL at 57a3f86: inconsistent derived cells,
historical/current contradictions, and missing namespace, immutable repricing proof, validator
negatives and source derivation. Sol attempt 2 retains 81.5/129.5/51/69 and corrects those findings.
The minimal operator-preflight namespace arm is an explicit scope amendment needed to preflight the
new runner namespace; it changes no preflight behavior. Durable evidence is under
artifacts/issue368-floor-recount/. Full workspace and timed gates remain root-owned and unrun.

Focused outcomes: historical repricing/equality/negative proof PASS (46 records; source SHA-256
9bb03dbbfa33e502fe05c0724bbfe6bf45d704236e0c85177b3ef1d02961ebcb); cargo test -p bench PASS
(30 passed); console validator mutation suite PASS with 0 workload launches; operator preflight
PASS with 0 workload launches, one warmup configured and two measured rounds configured. The
preflight also exposed and this tranche corrected its pre-existing repository-root resolution
(scripts/operator needs ../..). Logs: /tmp/engine-368-repricing.log,
/tmp/engine-368-bench-test.log, /tmp/engine-368-validator.log, /tmp/engine-368-preflight.out and
/tmp/engine-368-preflight.err.

Attempt 3 evidence record (2026-09-04)

Astra's attempt-2 verdict was FAIL at 5a156c46: the proof invoked only floor_shape, current
full validators correctly rejected historical .toml schemas, and bare inverted jq commands could
mask acceptance or execution errors. The synchronized applicability amendment authorized this
final correction. Historical revision dc581f3470b40678301d9504f1be4b1fd6be7173 supplies the exact
library and full validator pair that accept all 46 original mono3 records. A temporary library copy
changes only compressor_lane_ops 94 to 81.5 and limiter_lane_ops 138 to 129.5; its full record
and aggregate gates accept the exactly-three-derived-field repricing. Stale originals and an
unrelated measured-field mutation reject with exit 1; acceptance and broken-jq self-mutations prove
the negative harness distinguishes valid input and execution errors. Original/repriced library
SHA-256: 892e4e7c0810dba4f33121c00ceca8cd035f55119e38e2a433314b72d703baac /
3984c3d4825a0a426bd0172e67b07b63b756097fcd0cc0a0b57ac201e562acc1; record/aggregate validator
SHA-256: bfadf1b47e079246f86e7a6c676f466d48da720e9fbdd35a12b918e1dfc58162 /
d4160668c3947d7ee5d3b7522673342c64429e6cff47f87344ab71cfdc2556f5.

Focused proof PASS (/tmp/engine-368-attempt3-repricing.out); unchanged current-schema validator
suite PASS with zero workload launches (/tmp/engine-368-attempt3-current-validator.log). Root's
candidate workspace comparison at 5a156c46 passed 1546/0/24 versus baseline 1545/0/24, the one
additional pass being the focused floor test (/tmp/engine-368-workspace.log). No timed workload
has run; the fresh current-schema benchmark remains root-owned after Astra approval.

Spec-state finding

The local numbered-spec scan found no files for #366–#394, including #368's promised .github/ISSUE_SPECS/368-lane-13-floor-recount.md. Root reports all 172 existing numbered specs have matching remote issues. Before implementing #368, mirror its current remote body and this reconciled brief locally with exact issue number/title; synchronize the remote scope/role amendment. Do not manufacture new numbered issues for existing remote issues, edit other agents' active issue claims, or bundle unrelated historical spec restoration into this PR. Record remaining mirror debt as a bounded synchronization task.

Attempt 2 verdict and final-attempt evidence applicability amendment

#368 attempt 2 adversarial review — Astra

VERDICT: FAIL. Do not authorize the benchmark. Reviewed checkpoint 5a156c46 cumulatively against approved spec ed4b543b and attempt-1 findings at 57a3f86c. The document arithmetic/history corrections, minimal runner arm and bounded operator-root fix are acceptable. The remaining blocker is discriminating record-validation evidence, not the floor inventories.

Blocking evidence defects

1. verify-historical-repricing.sh invokes only floor_shape, not the actual full record and aggregate validators expressly required by the brief. I ran both real validators read-only on the committed repriced file; both exit 1. This is not a hypothetical omission:

jq -e -L scripts -f scripts/console-benchmark-record-validator.jq \
  artifacts/issue368-floor-recount/historical-mono3-repriced.jsonl
# exit 1
jq -s -e -L scripts -f scripts/console-benchmark-validator.jq \
  artifacts/issue368-floor-recount/historical-mono3-repriced.jsonl
# exit 1

A concrete independent incompatibility: mono3 retains .toml fixture IDs; current session_kind_shape requires .json (console-benchmark-record-lib.jq:184–208). A read-only inventory found no sealed accepted console stream whose first row has both real core_clock_hz and a .json fixture ID. Therefore the original brief's combination of full CURRENT validation and historical equality apart from exactly three floor fields is not satisfiable with the identified historical evidence. This fact needs an explicit spec correction, not field rewriting, validator weakening, or silently substituting floor_shape for the full gates.

2. Three standalone ! jq ... negative checks do not enforce rejection under set -e. bash -c 'set -e; ! true; echo reached' prints reached: an inverted command is exempt from errexit. The script can print PASS when a stale floor or deliberately malformed measured value is accepted. Use explicit if jq ...; then ... exit 1; fi checks and reject unexpected jq execution/parsing errors separately from a predicate false if claiming the failure class. This is a bounded local evidence-script correction, not generic gate infrastructure work.

Necessary scope decision before the final Sol attempt

Root has approved the following bounded applicability correction during this review; mirror it into #368's local spec and remote issue before attempt 3. The final Sol revision MUST:

• Keep the sealed originals and the exactly-three-field equality proof unchanged.
• Retrieve the historical record+aggregate validator pair and library from a clearly identified source revision that accepts the original mono3 records; prove unmodified originals pass it.
• In a temporary/evidence namespace only, substitute just the justified floor inventory constants in that historical validator library. Run the full historical record and aggregate gates on the repriced copy, and show stale originals plus unrelated invalid fields reject through explicit negative assertions. Retain minimal provenance/diff evidence so no semantic validator relaxation is hidden.
• Prove the negative assertion logic itself: feed an otherwise valid record into each supposed rejection path (or invert the intended mutation) in an isolated self-check and require the proof script to fail, never print PASS. A jq syntax/runtime error must be reported as execution failure, not accepted as the intended predicate rejection.
• Separately retain the existing current-validator positive/negative floor suite, and require the single fresh benchmark to pass the CURRENT full record/aggregate validators. Optional missing fresh cycle columns remain honest; historical full-gate proof is what then exercises measured cycle-bearing records.

This reconciliation is approved for the final Sol revision, conditional on root synchronizing the spec amendment first; it is not permission to modify current production validators or claim current-schema historical acceptance. If the historical validator capture cannot be identified and validated in one bounded correction, split/rescope the qualification remainder and stop instead of constructing a migration framework. One Sol attempt remains under the three-attempt limit.

Non-blocking/checkpoint disposition

The corrected eq+comp isolate percentages and ragged floor now calculate correctly; retrospective repricing is labeled and the mono/select debt wording is corrected. Fixed core throughput remains unchanged. The shared expression derivation is now recorded. The operator preflight's scripts/operator/../.. root is the correct repository path, and the namespace addition is minimal. No benchmark has run. Existing preflight evidence names old HEAD57a while exercising the working edits; root should rerun the non-timed preflight against the final committed head once the proof is fixed, so the final candidate identity is exact.

No cargo tests, benchmark, source edits, GitHub actions or spawned agents were used by this review. Full workspace is root-owned and pending; a workspace PASS cannot remedy the two evidence defects above. Actual PR still requires Astra review before merge.

Final local gate evidence (2026-09-04)

Astra attempt-3 checkpoint PASS at 6d64f11eded121761b5e835e7f98ce41dda6b892; full historical-schema record/aggregate proof and its assertion self-mutations pass. Current console validator suite, 30 bench tests, formatting, realtime policy and workspace clippy pass. Candidate workspace: 1546 passed, 0 failed, 24 ignored versus baseline 1545/0/24; +1 floor test.

The final committed-head preflight passed with zero workload launches. The sole runner invocation passed with one warmup, two measured rounds and 46 records accepted by the CURRENT full record/aggregate validators. Disposition is controlled (CPU63, quiet sibling, loadavg0.32); cycle fields honestly absent because hardware counters were unavailable. Historical measured cycle-bearing repricing is proven separately under its recorded full schema, with only the three authorized floor fields changed. Durable output, preflight metadata, provenance and limits are in artifacts/issue368-floor-recount/. No rendered arithmetic or runtime speedup claim changed.

Actual PR Astra review and required CI remain pending before merge.

[tamashi095]

Luna attempt-1 checkpoint 57a3f86c: accounting updates and focused tests are upstream. cargo test -p bench (30 tests), hermetic console validators (zero timing), realtime policy, formatting and diff checks pass. Workspace baseline 87926988: 1545 passed, 0 failed, 24 ignored. Dedicated runner namespace, historical repricing proof, candidate workspace comparison and benchmark remain pending; no timing invoked and no completion claimed. Astra is reviewing scope completeness before proceeding.

[tamashi095]

Astra adversarial review of Luna attempt 1 at 57a3f86c: FAIL. Core inventory counts are retained, but inconsistent repriced table cells, remaining historical/current contradictions, the missing dedicated runner arm, and missing immutable historical-record repricing proof require a bounded correction. Sol is assigned attempt 2 under the synchronized brief. No timed workload has run; no gates are weakened.

[tamashi095]

LANE-13 (#368) merged in 0e248bb07cfbf7dd136ec48649ec61ee9171d15b via PR #397 after Astra PASS at exact head adf936c3905d3e5a1546a96ef03eb40ae259fb15 and green required qualification. Current compressor/limiter inventories are 81.5/129.5 lane-ops (EQ51, builtins69; whole strip331), with Rust/jq authorities synchronized and historical evidence distinguished from current pricing. No rendered arithmetic changed. Luna attempt1 and Sol attempt2 failed review; Sol attempt3 passed the explicit historical-schema applicability correction and discriminating negative checks. Full historical record/aggregate proof preserves all but three floor-derived fields; current validator tests and the sole fresh run pass (one warmup, two rounds,46 accepted records; controlled CPU63, no hardware cycle fields). Workspace1546/0/24 vs baseline1545/0/24; one new floor test. All raw/accepted/disposition evidence is upstream. #368 verified CLOSED.

Current Wave0 coordination: #370 docs-only correction is also merged/closed via #395. #369 and #371 remain claimed/open with existing PRs; #306 must land after #371. The recorded Wave0 prerequisite therefore still blocks Wave1 implementation. Native AArch64 and the separate automation-drain feature retain their explicit owner deferrals.