Hoist mixed input-chain elision dispatch out of frame loops

[tamashi095]

Authority and smallest closable outcome

Parent: #559 RT7. Companion tracker: #560. Base and current main are 898bdc94b0143288049397629f3afeded384f8c2.

Sol HIGH coordinates. Astra XHIGH implements because this is low-level render-time SIMD kernel work. A separate Astra LOW reviewer performs scope review and each adversarial verdict. Agents are bounded assignees and do not own the issue.

#709 is passive at SOURCE PASS and consumes no implementation slot. Lane B independently owns #705 and every AudioWorklet qualification/pin decision. This issue may occupy lane A's one active implementation slot without waiting on disjoint lane-B work.

The smallest closable outcome selects each dual or mono input channel's two-section elision shape once before its frame loop, then runs a bounded specialized body whose loop neither reads InputChainPlan nor branches on plan elision. It preserves the exact per-channel arithmetic order, identity-run add(+0.0) placement, state writes, reports, and output bits for every one of the four two-section shapes at scalar, W4, and W8 widths. No DSP equation, coefficient, state format, public API/ABI, session schema, benchmark, artifact, pin, or performance budget changes.

Current fact pattern

input_chain_block_elided already dispatches the all-unelided and all-elided four-section cases once per block. Its mixed dual body, mixed_chain_block, still reads plan.elided[channel][section] and branches inside the frame/channel/section loop. The newer mono-collapse sibling mixed_chain_block_mono repeats the same plan read inside its frame/section loop. Both plans are invariant for the call: HPF/LPF coefficients are prepared-only, and reset/state restoration recomputes the plan.

The existing elision corpus proves bit identity across widths and section patterns, including signed-zero placement and retained state. It does not by itself prove plan selection moved outside the frame loops. The audit's location and cost are structural hypotheses, not current timing evidence. No timing or speedup claim is authorized.

Exact-path ownership

Implementation and private/unit evidence:

crates/lane/src/kernels/builtins.rs

Existing integration evidence may be extended only at:

crates/lane/tests/input_chain_elision.rs

Decision record:

.github/ISSUE_SPECS/710-hoist-mixed-input-chain-elision-dispatch-out-of-frame-loops.md

Coordinator-only concise rows in #559/#560/#349 are synchronized outside this checkout. No other path is owned. If the correct change requires another crate, public contract, benchmark framework, generated artifact, policy change, or timing claim, stop and split it into a new stateless issue.

Product contract

1. Mixed dual and mono input-chain execution each select a channel's exact [elide_hpf, elide_lpf] shape once outside that channel's frame loop. A private match plus const-generic or equivalent bounded specialization is acceptable. No frame-loop body may read InputChainPlan, index plan.elided, or branch on a runtime elision flag.
2. All four per-channel shapes retain the current order:
   • neither elided: HPF then LPF;
   • HPF elided: one add(+0.0) before LPF;
   • LPF elided: HPF then one add(+0.0);
   • both elided: exactly one add(+0.0) and no recurrence/state write.
     Sanitization, trim, nonfinite accumulation, store, and report order remain unchanged within each channel.
3. Dual left/right channels remain independent and use their own plan, coefficient, state, report, and buffer. Mono collapse continues to use channel 0 only. Processing one channel's complete block before the other is permitted only if the implementation and adversarial review establish that no shared mutable state, fallible step, observable side effect, or cross-channel arithmetic dependency changes behavior.
4. Output PCM, InputChainReport, and retained state are bit-identical to the current implementation for scalar, W4, and W8; all section patterns, signed-zero cases, nonfinite inputs, zero and short blocks, and representative multi-frame blocks are covered. Elided sections remain unwritten.
5. The plan remains prepared/recomputed at the existing boundaries. No live parameter, reset, restore, collapse eligibility, or symmetry rule changes.
6. Render stays allocation/free, lock, I/O, logging, syscall, panic-edge, and structural-mutation free. Work remains statically bounded by width and frames.
7. No public API or ABI expansion is required. Existing callers and InputChainPlan layout/meaning remain unchanged.

Attempt and checkpoint discipline

The repository's five-attempt maximum applies. Each attempt is one coherent Astra XHIGH implementation pass followed by a separate Astra LOW adversarial verdict. A failed prerequisite, compile, bit-identity, state/report, structural, policy, portability, allocation, or evidence gate stops that attempt. No retry within an attempt and no weakened gate. After five failures, preserve evidence and hard-stop; no disguised sixth attempt.

Astra XHIGH edits only the owned paths. It stops at each coherent compiling/focused-test checkpoint for Sol to audit, commit, and push before more implementation or qualification is layered on. Raw targets, compiler streams, mutation copies, binaries, and temporary evidence remain outside Git.

Objective gates

Before implementation, Astra LOW must verify exact main/branch/spec/GitHub identity, clean exact-path ownership, slot independence from #705/#709, the current dual and mono inner-loop plan-read shape, and that the proposed specialization can preserve the four operation orders without a public contract change.

Each implementation attempt must provide:

1. Focused scalar/W4/W8 tests over every dual left/right plan pairing and every mono plan shape, comparing output words, report words, and state words bit-for-bit with a frozen reference copy of the pre-change mixed bodies. Include signed zero, nonfinite sanitization, zero/one/multi-frame blocks, and elided-state non-write assertions.
2. A discriminating structural gate proving plan selection occurs once per processed channel per call and no runtime plan read/branch remains inside either mixed frame loop. An external mutation restoring an inner-loop plan lookup must be rejected for the intended reason. Compiler/lowering evidence may supplement, but not replace, a gate tied to this source claim.
3. Existing lane input-chain elision, sanitization, mono-collapse, and relevant builtins integration tests.
4. A render allocation gate covering mixed dual and mono shapes.
5. cargo test --locked -p lane in debug and release-unwind, strict lane Clippy, formatting, and exact diff/path review.
6. Workspace and realtime policy/mutation gates plus supported native x86-64-v3 and Wasm scalar/simd128 compilation appropriate to lane.

Astra LOW independently reviews the exact pushed checkpoint and reruns the claim-discriminating focused/structural gates before recording SOURCE PASS or FAIL. SOURCE PASS grants no timing, artifact, pin, PR, merge, or delivery claim.

Delivery boundary

After SOURCE PASS, Sol performs exact-head/current-main review. lane is in the browser artifact dependency closure, so lane B alone owns any separately numbered AudioWorklet applicability/qualification and pin decision. Lane A does not run a builder, qualify an artifact, or change a pin, and a passive source issue consumes no implementation slot.

Required PR qualification, guarded live head/base merge, post-main qualification, exact GitHub synchronization, and clean delivered-worktree removal follow the repository workflow. All failed branches, worktrees, commits, mutation copies, and evidence directories are preserved.

Attempt 1 — FAIL

Astra XHIGH supplied an uncommitted implementation from clean pushed brief
413767be8f3a54d47b33296bdf767772aca7da60; base remains
898bdc94b0143288049397629f3afeded384f8c2. External evidence is
/tmp/issue710-attempt1-xwq9EPTa. The original product source and extracted
mixed bodies are preserved there; original builtins.rs SHA-256 is
365fe127ef70d579bc35eb3817d57b62bf2d9a303908d37d952aef3c0a4e90fb.

Each mixed wrapper selects each processed channel through one four-arm match,
then invokes a const-generic frame body with no runtime plan parameter. Its
three const conditions retain HPF→LPF order and exactly one add(+0.0) at the
original identity-run position. Dual channels have separate buffers,
coefficients, local integrators and report accumulators; the existing scalar,
W4 and W8 lane primitives and recurrence have no cross-channel dependency or
observable side effect. Caller state is still published once after both
channels complete. Mono publishes only channel 0 and duplicates its report.
No public contract, prepared-plan boundary or arithmetic primitive changed.

The integration evidence retains the actual pre-change mixed dual and mono
bodies, renamed only, and reuses the existing designs, signal corpus and
output/state/report packer. Supplemental cases cover all 16 channel-plan
pairings, two section orders and 0/1/17 frames at scalar/W4/W8, with different
channel trim/data/state, signed zero, nonfinite inputs and forced elided-state
sentinels. Mono explicitly checks channel-1 and elided-section state remains
untouched. A private counter checks one selection per processed channel
independent of frame count. One production structural validator checks the
wrappers, four-arm selector, const frame conditions and the existing
no_lanes/svf_step/flush helper closure; no frozen reference participates
in that structural assertion.

The attempt stopped at its first failed gate without a retry. Commands ran
from /home/bl/misofm/engine-rt7-hoist-elision-plan, with Cargo output under
CARGO_TARGET_DIR=/tmp/issue710-attempt1-xwq9EPTa/target:

Gate	Result	Log
cargo fmt --check	Exit 0	01-fmt.log
Compare both frozen bodies with the actual pre-change bodies, restoring only their names	Exact equality	02-frozen-reference.log
cargo test --locked -p lane --test input_chain_elision -- --nocapture	Exit 0; all 9 tests passed, including scalar/W4/W8 bitwise and production structural checks	03-input-chain-elision.log
cargo test --locked -p lane --lib mixed_elision -- --nocapture	Exit 101; unqualified thread_local! and both vec! calls are unavailable in the no_std lib-test context; unresolved MIXED_PLAN_SELECTIONS errors follow	04-selection-count.log

Selector instrumentation did not compile. Full lane tests, release-unwind,
Clippy, allocation, external structural mutation, policy and portability
qualification were not run. The failed uncommitted state is preserved as
failed-attempt.diff and failed-state-sha256.json under the Attempt 1 evidence
directory. SHA-256 values are:

Preserved input	SHA-256
Product builtins.rs	402c3b38bb7e610c5921a22cffa53c40801eb24a9b2dcc32d5e667a6b7660f8d
Integration input_chain_elision.rs	b30e02d8171ae994ee53a5f09d6088d7f141745e3a5dd9a33c1a7008d805023b
Issue spec before this failure record	ec9b55b8a73a5f8a0175c6e4537fdf2f0b5a6831f407904cf22e8cfcdee13353
failed-attempt.diff	a60a60281da7f723a94b0519f0c980cfe71f4e97312d9cfc7c068757e3ec8b67

Astra LOW (/root/issue709_astra_low_scope) recorded Attempt 1 FAIL on that
inspected state: the lib-test compile failure is blocking, despite nine passing
integration tests. No additional blocking production finding emerged. The
review found the const shapes and independent channel execution preserve the
specified order, mono leaves channel 1 untouched, and the approximately 374
integration-test lines are proportional. Selector instrumentation and broader
qualification remain incomplete; no SOURCE PASS was recorded.

Attempt 2 — required implementation gates complete; review pending

Astra LOW and Sol authorized only qualifying the test macros as
std::thread_local! and std::vec!, updating this record and rerunning the
complete sequence. No suppression, production change or additional path is
authorized. Astra XHIGH applied those three qualifications without changing
the integration tests. Attempt 2 evidence is separate at
/tmp/issue710-attempt2-IYCU7pK6; Attempt 1 evidence remains preserved.

The correction's formatting, frozen-reference provenance, nine integration
tests and one selector-count test passed (01–04 logs and
focused-results.jsonl). Astra XHIGH paused; Sol audited, committed and pushed
the exact source checkpoint c5b08616454d7367fc3e758cd541a0873c503ba5.
Sol then authorized a restart of the full sequence on that frozen source.
All commands below passed without a failed gate or retry. This record is an
implementation evidence report, pending the separate Astra LOW verdict.

Commands ran from /home/bl/misofm/engine-rt7-hoist-elision-plan with Rust/Cargo
1.97.1. Let E=/tmp/issue710-attempt2-IYCU7pK6; the default Cargo target was
CARGO_TARGET_DIR=$E/target. Native commands inherited the repository's
approved -C target-feature=+avx2,+fma configuration. Policy commands also
used TMPDIR=$E/policy-tmp; scripts with explicit /tmp scratch paths kept
their existing behavior. toolchain-and-environment.log records toolchain,
installed targets and flag environment. full-results.jsonl records literal
argv, environment overrides, source commit and exit for every row. Log names
below are relative to E and have the suffix .log.

Exact command (environment overrides shown where different)	Result	Log
cargo fmt --check	Exit 0	full-01-fmt
python3 "$E/verify-frozen-references.py"	Exit 0; both actual pre-change bodies equal their frozen copies after restoring only function names	full-02-frozen-reference
cargo test --locked -p lane --test input_chain_elision -- --nocapture	Exit 0; 9 passed	full-03-input-chain-elision
cargo test --locked -p lane --lib mixed_elision -- --nocapture	Exit 0; 1 passed	full-04-selection-count
CARGO_TARGET_DIR="$E/mutation-target" cargo test --offline --manifest-path "$E/inner-plan-mutation/Cargo.toml" --test input_chain_elision mixed_elision_production_structure_selects_before_frames_and_audits_helpers -- --exact --nocapture	Compiled; expected test exit 101 at the intended structural assertion; validator exit 0	full-05-inner-plan-mutation
cargo test --locked -p lane --test sanitise_counter	Exit 0; 2 passed	full-06-sanitise
cargo test --locked -p builtins --test mono_collapse --test input_liveness_mono	Exit 0; 10 passed	full-07-builtin-mono
CARGO_TARGET_DIR="$E/allocation-target" cargo run --offline --release --manifest-path "$E/allocation-harness/Cargo.toml"	Exit 0; detector liveness and 288 audited calls passed	full-08-allocation
cargo test --locked -p lane	Exit 0; 48 passed, 2 descriptive timing tests ignored	full-09-debug
CARGO_PROFILE_RELEASE_PANIC=unwind cargo test --locked -p lane --release	Exit 0; 48 passed, same 2 ignored	full-10-release
cargo clippy --locked -p lane --all-targets -- -D warnings	Exit 0	full-11-clippy
bash scripts/check-workspace-policy.sh	Exit 0	full-12-workspace-policy
bash scripts/check-lane-policy.sh	Exit 0	full-13-lane-policy
bash scripts/check-realtime-policy.sh	Exit 0; 41 regions in 12 files	full-14-realtime-policy
bash scripts/check-unfused-seal.sh	Exit 0	full-15-unfused-seal
bash scripts/check-realtime-audit-leak.sh	Exit 0; production dependency graphs exclude audit instrumentation	full-16-realtime-audit-leak
bash scripts/test-workspace-policy.sh	Exit 0; expected counter-mutation diagnostics retained	full-17-workspace-mutations
bash scripts/test-lane-policy.sh	Exit 0	full-18-lane-mutations
bash scripts/test-realtime-policy.sh	Exit 0	full-19-realtime-mutations
bash scripts/test-realtime-audit-leak.sh	Exit 0	full-20-audit-leak-mutations
cargo build --locked --release -p lane --target x86_64-unknown-linux-gnu	Exit 0; approved native ISA flags	full-21-native-build
CARGO_TARGET_DIR="$E/wasm-scalar" cargo build --locked --release -p lane --target wasm32-unknown-unknown	Exit 0	full-22-wasm-scalar
CARGO_TARGET_DIR="$E/wasm-simd128" CARGO_TARGET_WASM32_UNKNOWN_UNKNOWN_RUSTFLAGS="-C target-feature=+simd128" cargo build --locked --release -p lane --target wasm32-unknown-unknown	Exit 0	full-23-wasm-simd128
python3 "$E/verify-final-source.py"	Exit 0; exact HEAD/upstream, clean tree, base/brief diffs confined to the three owned paths, unchanged product/test hashes, actual brief provenance and prior gate statuses verified	full-24-source-review

The one external mutation passed the runtime plan into the specialized body
and restored all three per-frame conditions to plan.elided lookups. The
production structural test rejected it with
runtime plan/control in frame body or helper: mixed_channel_block: InputChainPlan.
It compiled and reached that assertion: 0 passed, 1 failed, 8 filtered out.
The isolated manifest emitted two existing miso_wasm_simd8 check-cfg warnings
because it does not inherit the workspace lint declarations; these did not
cause the rejection. The input copy, manifest, source, test and exact
mutation.diff remain under E/inner-plan-mutation; their hashes are in
mutation-input-sha256.json. Mutation-diff SHA-256 is
63f7ceeac7d6053148dbe5679edc0ddb408bd07d1ea18a9de6b10b028a35bdb5.
The authoritative checkout was never mutated for this gate.

The external allocation harness used the public dual/mono elided entries,
actual prepared plans for all 16 pairings, different channel data/trim,
nonfinite and signed-zero inputs, and 0/1/17 frames at scalar/W4/W8. Its
288 scoped calls include every mixed shape and the existing fast paths;
each had zero allocations and frees. Mono channel-1 sentinels stayed intact.
The System-forwarding detector first observed its expected 3 allocations and
3 frees across alloc, zeroed alloc, realloc and dealloc. All buffers were
prepared outside the measured intervals; this checks kernel allocation, not
whole-host rendering or syscalls. Harness source, manifest, generated lockfile,
binary and target remain external. allocation-input-sha256.json records
manifest SHA-256 330d8145d95b0653aba6b26e4a2abc1505b52d5686c4fd27e85297aaa634c393
and source SHA-256 0fa81daf3d655a7f265af1b972d3e70b502bd0ecad62ba8be88573e6347c7810.

qualified-source.diff preserves the reviewed base-to-source diff.
frozen-checkpoint-sha256.json and the final review confirm product SHA-256
217815ba5ddddf85a3a3ed7d522103a8a8b20730e69f395fcdc2a29ce42e08f1
and unchanged integration SHA-256
b30e02d8171ae994ee53a5f09d6088d7f141745e3a5dd9a33c1a7008d805023b.
No source or test edit followed the checkpoint; only this decision record
changed after qualification.

Bitwise and allocation executions are native scalar/W4/W8 evidence. Wasm
results establish compilation only; no Wasm runtime or AArch64 execution was
performed. The structural gate checks the production source and its named
free-helper closure, not an assembly or cycle claim. No timing, speedup,
AudioWorklet, artifact, pin, PR, merge, delivery or SOURCE PASS claim is made.
Astra XHIGH pauses here for Sol's exact-path checkpoint audit and Astra LOW's
independent adversarial review.

Attempt 2 — Astra LOW SOURCE PASS

Astra LOW (/root/issue709_astra_low_scope) returned SOURCE PASS at exact
clean pushed head a1890963eb383e5aad1985a1f0efb8a4f3140fb4 for product/test
checkpoint c5b08616454d7367fc3e758cd541a0873c503ba5. The reviewer verified
HEAD/upstream identity, local/GitHub body parity, exact three-path ownership,
and all preserved input, mutation and evidence hashes. Fresh review evidence is
at /tmp/issue710-low-review-lzqkanjf.

The reviewer independently reran:

Command	Result
python3 /tmp/issue710-attempt2-IYCU7pK6/verify-frozen-references.py	Exit 0; original source also matched the Git base
cargo test --locked -p lane --test input_chain_elision -- --nocapture	Exit 0; 9 passed
cargo test --locked -p lane --lib mixed_elision -- --nocapture	Exit 0; 1 passed
cargo clippy --locked -p lane --all-targets -- -D warnings	Exit 0
cargo run --offline --release --manifest-path /tmp/issue710-attempt2-IYCU7pK6/allocation-harness/Cargo.toml	Exit 0; detector liveness and 288 allocation/free-free calls passed

No blocker was found. Arithmetic and signed-zero placement remain intact;
disjoint channel execution preserves reports and state, mono preserves channel
1, and instrumentation is excluded from production by cfg(test). Frozen
references, selector counts, source/helper checks and the specifically rejected
production mutation provide complementary evidence. The structural gate covers
the current named helper closure rather than arbitrary transformations or
machine code; allocation evidence covers the tested kernels; Wasm evidence is
compilation only.

Sol fetched origin after the verdict. origin/main and the branch merge base
remain 898bdc94b0143288049397629f3afeded384f8c2; there is no current-main
drift and the diff remains limited to the three owned paths. lane is in the
browser artifact dependency closure, so #710 is passive pending a lane-B-only
artifact disposition and consumes no implementation slot. No timing, speedup,
AudioWorklet, artifact, pin, PR, merge or delivery claim is granted.

Current-main integration and lane-B delivery scope

User authorized takeover with Astra LOW implementation and Astra XHIGH
verification. Independent XHIGH scope review revalidated all 38 preserved
evidence hashes, two allocation-input hashes and five mutation-input hashes.
The full lane/builtins source, Cargo manifests/lock, toolchain and native flags
are unchanged between original base 898bdc9 and current main f991f8a.
The accepted implementation is merged unchanged onto f991f8a, preserving
92a00bc history and the exact three-path feature diff.

Current-base focused qualification passed once: formatting; nine lane elision
integration tests in debug and release-unwind; one selector-count unit test;
and ten builtins mono/liveness tests. Commands, actual exits and logs are
external at /tmp/issue710-delivery-ZJL89IA7. Original discriminating mutation
and 288-call allocation evidence remain applicable; no timing claim follows.
This is delivery integration, not a new product implementation attempt.

Lane B root additionally owns only these delivery paths for this issue:

• hosts/host-web/web/miso-engine-v1-audio-worklet-artifact.sha256
• hosts/host-web/qualification/results.json
• hosts/host-web/BROWSER_DEPLOYMENT_MATRIX.md

Run the existing builder's repin-report mode once on the integrated source in
fresh external directories. If its digest differs from current pin
5695fbc4d72fae4a78b5acd1cf8970c489163703a11ac5351974ce05a90b1574,
root may checkpoint that digest as provisional and unqualified. Then the
ordinary builder must produce the exact six-file artifact in a fresh empty
external directory. Qualify that single artifact through existing static,
resource/native parity, hermetic host, SDK type/headless/package and three-
browser gates. Record the matrix only from successful actual browser results
using the integrated source commit and existing --record-matrix mechanism.
Non-Wasm files must match current main (including #719 host JS/declarations);
browser floors, gates and resource expectations must not change. Unexpected
drift or failure stops for attribution; no policy/fixture weakening, new
qualification framework, benchmark, generated compiler capture or publication.

After XHIGH exact-source/artifact review, deliver through exact-head required
PR qualification, guarded merge and successful exact-main qualification.
Synchronize issue closure only then; remove clean delivered worktrees while
preserving branches/history and external evidence. #709 and #714 remain outside
this integration.

Provisional artifact checkpoint

Integrated source 5a66a1fa41e57437cfc76ee27d30a59fce0dbbcb produced digest
6d3a6144109bbc4578dd661bdb84a13490fc94c5191b72608409ca40b44581f8 in the single authorized repin-report invocation (exit 0).
Root records this as a provisional, unqualified pin so the ordinary builder can
reproduce and qualify it. No artifact acceptance, SOURCE delivery or browser
claim follows. Exact command/environment/stdout/stderr are external at
/tmp/issue710-artifact-88js3ptr/probe.*; locked SDK/browser dependency setup
also passed. The next step is one ordinary six-file build and existing gates.

Current-source and artifact qualification checkpoint

Independent Astra XHIGH records INTEGRATION SOURCE PASS for 5a66a1f: exact
merge parents f991f8a and 92a00bc, accepted product/test hashes unchanged,
all five focused gate receipts verified. Subsequent c375ca9 changes only the
provisional pin/spec.

The ordinary builder reproduced the observed digest and exact six-file set.
Every following command exited 0 against that same external artifact (no
retries). Let A=/tmp/issue710-artifact-88js3ptr/artifact:

bash scripts/build-web-audioworklet.sh "$A"
bash scripts/check-web-audioworklet.sh "$A"
python3 -B scripts/check-browser-expected-resources.py --artifacts "$A"
bash scripts/test-web-audioworklet.sh
bash scripts/check-sdk-types.sh
bash scripts/check-sdk-headless.sh "$A"
bash scripts/sdk-package.sh check "$A"
npm --prefix hosts/host-web/qualification run qualify -- --artifacts "$A" --browser all --record-matrix --candidate-commit 5a66a1fa41e57437cfc76ee27d30a59fce0dbbcb --self-test-mutations
node hosts/host-web/qualification/generate-matrix.mjs --check

Ordinary build ran with repin mode unset. Actual argv, source commit, external
TMPDIR/CARGO_TARGET_DIR, timestamps, exits, stdout and stderr are preserved at
/tmp/issue710-artifact-88js3ptr/, with qualification-commands.json and
artifact-sha256.json. SDK headless passed 188/188; publishable-tarball passed.
Static render closure, boot budget, native/browser corpus/resource parity and
26 resource negative controls passed. Chromium 151.0.7922.34, Firefox 153.0 and
WebKit 26.5 passed every existing matrix gate. Generated results differ from
main only in source/digest lineage; floors and expectations are unchanged.
All five non-Wasm payload identities remain current-main identities, including
#719 host JS/declarations; only the Wasm digest changes. These are existing
headless browser correctness gates, not physical-device or performance claims.

Final XHIGH artifact/evidence review and exact-head PR/main qualification are
still required before synchronized closure. No registry publication is part of
this delivery.

Astra XHIGH final artifact/evidence PASS

Independent review accepted exact pushed 2abc1cf87b780086075544baa09b8804d85a38f4:
source integration and accepted product/test identities, all six actual payload
hashes, probe/ordinary/pin/matrix agreement, all command receipts and successful
gates, unchanged five non-Wasm identities and lineage-only matrix differences.
No blocker and no rerun. This appended verdict changes only the decision record;
exact-head PR qualification, guarded merge and post-main qualification remain
required before closure.

Delivered and synchronized — 2026-09-10

PR #721 merged reviewed head ac53495 as
878db25. Guarded merge parents are exact
f991f8a and ac53495. Required PR qualification 34485013622 and exact-main
qualification 34485792089 both passed through their aggregate qualification
verdicts. RT7's bounded plan-dispatch removal is delivered with the qualified
6d3a6144 artifact. No speedup or physical-device claim is made. This evidence
checkpoint is upstream before issue closure; original source/qualification
logs remain external and branches/history are retained during worktree cleanup.

[tamashi095]

Astra LOW returned SCOPE PASS at clean pushed brief 413767be8f3a54d47b33296bdf767772aca7da60, base/main 898bdc94. It confirmed dual and mono form one bounded slice and can preserve behavior without a public contract change. Attempt 1 is assigned to Astra XHIGH in the three owned paths. Frozen pre-change references, production structural validation, differing dual-channel shapes/data, and untouched mono channel-1 state are required. #709 is passive SOURCE PASS; #705/#710 are the two independent active implementation slots. No timing, artifact, pin, PR, merge, or delivery claim is made.

[tamashi095]

Astra LOW verdict: Attempt 1 FAIL on the preserved uncommitted state based on 413767be8f3a54d47b33296bdf767772aca7da60. Formatting, frozen-reference provenance, and all 9 integration tests passed; the next lib-test gate failed to compile because unqualified thread_local! and vec! macros are unavailable in the no_std test context. Broader gates did not run. No production blocker was found. The exact failed diff/hashes and logs are preserved in /tmp/issue710-attempt1-xwq9EPTa. Attempt 2 may only qualify those macros, update the record, and restart the complete sequence; no production revision or suppression is authorized.

[tamashi095]

Attempt 2 buildable source checkpoint is pushed at c5b08616454d7367fc3e758cd541a0873c503ba5. Formatting, exact frozen-reference verification, all 9 input-chain elision integration tests, and the private selector-count test pass. Attempt 1 remains FAIL with its exact uncommitted diff/logs preserved externally. The full Attempt 2 qualification is pending from this frozen source; no SOURCE PASS, timing, artifact, pin, PR, merge, or delivery claim is made.

[tamashi095]

Astra LOW independent verdict: SOURCE PASS — Attempt 2 at exact clean pushed head a1890963eb383e5aad1985a1f0efb8a4f3140fb4, recorded at evidence head 92a00bc504436ec70e6c88a0909aaf9c03e38461. Fresh frozen-reference, 9-test bitwise/structural, selector-count, strict Clippy, and 288-call allocation reruns passed. Current main remains base 898bdc94. #710 is passive pending lane-B-only artifact disposition and consumes no implementation slot. No timing, artifact, pin, PR, merge, or delivery claim is made.

[tamashi095]

Delivered via PR #721 at 878db25. Independent source/artifact PASS; required PR qualification 34485013622 and exact-main qualification 34485792089 both PASS. RT7 dispatch removal is delivered; no speedup claim.