You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
code-index: a one-file edit re-verifies the whole sealed generation (~5 min, 2.3 GB peak) #4507
Measured by the daemon-memory investigation (PR #4503 lane) on a release build of origin/master 9b8789c, isolated profile, fresh clone of this repository (6.1k files):
A one-file edit triggers an incremental pass whose physical footprint peaks at 2.3 GB and that takes about 5 minutes; footprint returns to 0.38 GB afterwards.
The pass re-reads every sealed segment of the active generation instead of only what the changed file touches.
On the live daemon the same checkout logged refresh passes of 870 s and 1,489 s that ended outcome="noop".
Expected
An incremental pass costs in proportion to the changed files (re-extract the file, re-link its edges, reuse unchanged sealed segments), not a full-generation re-verify.
Why it matters
Every edit in an active checkout pays minutes of multi-core CPU and GBs of transient memory, which is a large share of the sustained daemon CPU users see while agents edit code.
Issue #4507 ownership update (2026-11-07): the principal sealed-generation proof reuse fix was already merged via #4523. I opened draft #4791 against master at fc4322c74abfd7d2beee88c8f70278c45deef123 to add a falsifiable real one-file-successor proof regression (cold versus parent-warm check counts, plus existing corruption refusal). This is test coverage, not a second copy of the production fix. Separately, #4520's sparse-successor CPU/RSS residual has active draft #4789; I have not modified that overlapping lane. Hosted Bazel CI and independent review are still pending; no native macOS runner/local Bazel is available in this run. Keep #4507 open until full end-to-end time/memory and corruption acceptance is demonstrated on master.
Incremental #4507 engineering update (2026-11-07): PR #4791 now has commit e16aaf0. In addition to cold-vs-parent-warm nonzero inherited-segment proof checks on a real one-file successor, it now warm-proves the successor, mutates its newly changed sealed file segment without changing length, requires typed Corrupt, restores the bytes and verifies recovery. This closes the blind spot where a verifier could have passed the inherited-read count assertion while skipping all changed source. PR moved to ready-for-review solely to trigger standard Bazel Linux CI; the general PR sweep has an explicit do not merge hold on #4791 pending exact-head CI and independent review. CI: https://github.com/ScriptedAlchemy/tracedecay/actions/runs/42434575039 (queued at this check); native macOS not part of the default run. No releases, Kev/decision-model merges, or overlapping #4520 source edits. Issue remains open pending end-to-end memory/time acceptance on master.
Published a further production fix in the existing #4791, commit cd1e40aada9d38801877ea62b92520d1daaf4f0a: preserve the last authenticated sealed-source snapshot across interrupted and overlapping proofs, without exposing tentative hashes or allowing stale completion to replace a newer successful proof. Four real-file regressions cover that reuse and the corruption/root/extent/deletion boundaries; the existing real one-file successor proof regression is retained.
No duplicate PR or overlapping #4520 implementation was created. Branch update was an expected-head, non-force fast-forward from 216e726ef58b46ce6bcc4017e1ac225a51fcf8c1.
Exact-head hosted Bazel CI is queued/pending at https://github.com/ScriptedAlchemy/tracedecay/actions/runs/42442313648; native macOS is skipped. No local Bazel is available, and no Cargo was run. Nonzero passing regression results, independent review and fresh full-pass time/RSS evidence are not yet established. #4791 remains under an explicit merge hold; this issue remains open. Other sweeps should coordinate on that PR rather than duplicating the fix or treating partial evidence as acceptance.
#4507 exact-head verification evidence (2026-11-08 UTC): Downloaded and inspected the actual hosted Bazel test artifact for existing PR #4791 at cd1e40aada9d38801877ea62b92520d1daaf4f0a. //crates/tracedecay-store-runtime:unit_test passed 177 tests, 0 failed, 2 ignored, 0 filtered, including all four newly added real-file verification tests and inherited_segment_proof_rejects_wrong_manifest_extent for the real one-file successor. This is a non-vacuous execution result, not just a green build. Test artifact: https://github.com/ScriptedAlchemy/tracedecay/actions/runs/42442313648/artifacts/13606073424.
New blocker: Overall CI #17306 is red despite successful repository formatting, Clippy, feature gates, dashboard, and the store-runtime tests. The three failing Linux test targets are automation-runtime (2 timeout assertions), runtime acceptance (LSP deadline), and automation-runner (3 scheduler/backend assertions). The latter three individual automation-runner failures are also present on unmodified master (CI #42328697300). The other failures are outside this PR's changed paths, but require separate baseline/flakiness attribution—no test weakening or bypass. Windows was still running, native macOS skipped, and independent review plus fresh real-workload full-pass time/peak-RSS measurements remain missing.
Updated the existing #4791 description with this evidence and kept its explicit merge hold. No second PR, code rewrite, Cargo execution, release, Kev/decision-model merge, or issue closure. Refs #4507 remains appropriate; not yet Fixes.
#4507 production-regression follow-up published to existing PR #4791: cb58c6d. Safely fast-forwarded from exact old head cd1e40a with force=false; no duplicate PR or unrelated source change.
New falsifiable real-path coverage: The prior one-file successor test checked that corruption of a changed sealed segment is refused and restoration succeeds, but did not prove that failed verification preserves authenticated unchanged source. It now counts production proof callbacks after the corrupted changed segment is restored: recovery must do a nonzero new proof, and on Unix a cold fresh verifier must perform at least one extra proof check per inherited file segment. This fails if the last good proof is evicted on error, so source integrity and incremental recovery cost are asserted together. Non-Unix still requires full proof.
Previous exact-head (cd1e40a) hosted Bazel artifact had 177 passing store-runtime tests including 5 #4507 regressions; that is not validation of new cb58c6d. New head CI #17387 is pending. Earlier overall Linux CI #17306 had unrelated automation/runtime acceptance failures; do not weaken or bypass them. Native macOS validation, independent review and fresh full-pass wall/peak-RSS measurements are still missing.
The merge hold remains, this issue remains OPEN with Refs #4507 (not Fixes), and the separate #4520/#4789, protected draft #4776 and read-only release #4763 are untouched. No Cargo, force update, sweep merge, release, or infrastructure action.
New exact-head CI blocker for #4507 (2026-11-08 09:04 UTC): PR #4791 run #17387 at cb58c6dd7e0806637eb63388ed229c67eb98046a passed repository gates, change-scope and benchmark-harness checks, but Clippy and Dashboard failed during Bazel dependency analysis because GitHub returned HTTP 504 fetching the pinned hermeticbuild/hermetic-launcherrunfiles-stub-x86_64-linux and finalize-stub-x86_64-linux assets. No compiler/test failure was reached for those jobs. Linux Bazel and feature gates remained queued. Native macOS was skipped.
I attempted to rerun the Clippy job directly, but GitHub refused with HTTP 403 because the containing workflow is still running. The retry must happen only when the run finishes; no workflow/infrastructure alterations or speculative source commits are warranted for this external 504. The PR description is updated with exact log-derived attribution. Its MERGE HOLD and Refs #4507 remain, with independent review, nonzero exact-head regression results and fresh full-pass wall/peak-RSS measurements still outstanding. Do not duplicate this issue PR, bypass checks or close #4507.
Additional exact-head #4507 CI blocker (2026-11-08 09:40 UTC):#4791 run #17387, head cb58c6dd7e0806637eb63388ed229c67eb98046a, also lost its Linux fleet Bazel build/test lane before tests could execute. Linux fleet job 128671373185 reports the same pinned hermeticbuild/hermetic-launcher GitHub release-asset HTTP 504 Gateway Time-out, now for both finalize-stub-x86_64-linux and runfiles-stub-x86_64-linux. Its bazel build --config=ci //... ended in Bazel repository-loading errors; no exact-head store-runtime test result is available. The dependent Linux gate failed because its fleet lane failed; this is not an additional source-test failure. Repository gates, benchmark-harness checks, change-scope and feature gates passed; Clippy/Dashboard previously failed the same upstream 504. The Windows job was still running at the latest poll, so the workflow remains in progress and the failed-job retry is not yet eligible. Native macOS was skipped.
No speculative source or infrastructure adjustment is justified by a third-party fetch outage. Keep Refs #4507, the #4791 merge hold, and #4507 open. Retry failed jobs through GitHub's native workflow retry only after the run completes; retain nonzero regression, independent review and fresh wall-time/peak-RSS evidence requirements.
Published a falsifiability fix to the existing #4507 PR #4791: 25caf8d. The previous post-corruption assertion recovered_reads > 0 could pass on the manifest/lifecycle check callbacks even if no repaired source segment was rehashed. The actual test now records a fully warmed successful baseline, demands a strictly higher callback count when verifying the changed segment after same-length corruption and restoration, and asserts a subsequent successful proof returns to the warm baseline. Existing cold-vs-inherited-digest bounds, typed corruption refusal, and cross-platform fallback stay intact. One file changed, tests only, no weakening or production/infrastructure change.
This was a verified non-force fast-forward of test/4507-inherited-segment-proof from exact expected SHA cb58c6dd7e0806637eb63388ed229c67eb98046a; no competing #4507 PR was created. New exact-head hosted Bazel CI #17439 is pending, not a claimed pass. Previous head #17387 lost Linux/Clippy/Dashboard to pinned hermetic-launcher GitHub HTTP 504 fetch errors. Previous cd1e40a store-runtime executed 177 passing tests but does not validate 25caf8d. Prior Windows job for #17387 was still building, preventing native failed-job retry then; hosted macOS remains skipped.
MERGE HOLD: No independent review or real 6.6k-file full-pass wall-time/peak-RSS evidence yet. Keep #4507 open, Refs #4507 only. #4520/#4789 are separate, draft Kev/decision-model #4776 is protected, release #4763 read-only. No Cargo, force update, auto-merge, Kody, Copilot or external delegation.
#4507: published a cross-platform regression correction to existing PR #4791 (2026-11-08 UTC).
Commit 1e530bf fixes a newly identified contradiction in the real one-file successor test. The previous assertion required recovered_reads > warmed_reads on every platform. But SealedSegmentFingerprintV1::of deliberately returns None on non-Unix systems, so warm and recovered proofs both rehash all segments and must produce equal callback counts. The test now retains the strict positive recovery-vs-warm proof delta on Unix (where caching exists) and asserts equal full-read costs on unsupported platforms. Typed corrupted-segment refusal, cold-vs-inherited nonzero bounds, and successful rewarming assertions remain unchanged. No production behavior, weakening of Unix acceptance, CI/infrastructure, or dependency changes.
Pushed an exact-parent, non-force fast-forward from 25caf8d700236e7bcf0ee07515f3a22bf5a7ade1 to 1e530bf2aedabbc8caaf5dcca5d24a85eb015ccf on the existingtest/4507-inherited-segment-proof branch; verified the new remote ref and one-file commit. Exact-head hosted Bazel CI #17446 is queued; no test pass is claimed for this head. Local Bazel/Bazelisk is unavailable; Cargo was not run.
MERGE HOLD stays. Need nonzero exact-head regression execution, independent review, and fresh real-workload wall-time/peak-RSS; native macOS remains skipped by the default workflow. Keep Refs #4507, issue open, no duplicate PR. #4520/#4789, protected draft #4776, and release #4763 remain outside this change.
Published issue-owned benchmark measurement fix a6bbbbbb6d47 on existing PR #4846 (Refs #4507). The existing scripts/bench-hot-paths.py already exercises a real disposable Git clone and daemon through an appended one-file edit, sync, and a new ready generation, with wall time and daemon peak RSS. But it previously started its wall timer and /proc RSS sampler after writing and closing the edited source file. Filesystem-watch-triggered refresh can therefore start before observation, systematically undercounting the very incremental work #4507 needs measured.
The patch moves timer and sampler activation ahead of the source mutation and keeps stop after the observed new generation is ready; sampling is stopped in finally on error. Added nonvacuous test_edit_timer_and_rss_observer_cover_the_file_mutation that drives the actual harness edit method against a tracked real file with a deliberately slow append, asserts sampling begins before the edit and stops after generation readiness, requires measured wall time to include the delay, and checks a positive peak-RSS result. Existing full fake-daemon end-to-end harness tests are preserved. No production graph/cache source change, deadline increase, weakened assertion or extra test-only production path.
Published safely via exact expected old head 85c8cd0, nonforce; independently verified PR + named ref at a6bbbbbb6d4740ea7949ff80b8888918ca973ccf, compare ahead 1 / behind 0, exactly two script/test files. Exact-head CI #18355 is queued, not passed. This is an accuracy fix for obtaining future same-corpus before/after evidence, not itself a representative performance measurement or proof that source rereads are eliminated. Rust Bazel tests/native validation, original ~6.1k-file same-host wall/peak-RSS comparison, and independent approving review remain outstanding. Keep #4846 unmerged and #4507 open; do not duplicate this work in sweeps. No Cargo, external delegation, protected model/release changes, production or infrastructure work.
Published e4064bbcb350 on existing #4846 (Refs #4507), expected-head non-force fast-forward from a6bbbbbb6d47. Independent PR and branch readbacks agree on the new SHA; compare confirms exactly one commit ahead / zero behind and two files, scripts/bench-hot-paths.py plus scripts/test-bench-hot-paths.py.
Actual measurement-correctness defect: the one-file benchmark waits for a generation different from the ID saved at initial indexing, but the daemon can complete another background refresh before the measured edit. The pre-existing generation ID then becomes stale; the old implementation could accept that unrelated already-completed refresh as the edit's result, reporting an artificially short pass and missing incremental peak RSS. A response claiming current/ready with an absent generation ID could also satisfy the old None != old_id check.
Repair: require a nonempty, current/ready generation ID immediately before mutating the tracked source, use that freshly observed ID as the baseline for the edit-to-ready wait, require a nonempty successor ID, and update the tracked generation after the measured pass. Existing timer/RSS observer still starts before the file mutation and stops after readiness, preserving the previous fix.
Falsifiable tests: the existing real tracked-file slow-edit regression now supplies an obsolete initial generation and an independently newer pre-edit generation; it asserts the wait uses the newer ID, includes the deliberate edit latency, and records peak RSS. It also requires a missing pre-edit generation to refuse without changing the source or starting sync, and a missing post-edit generation to be skipped until a later valid successor arrives (two status polls). Previous full fake-daemon harness tests remain unchanged.
Validation: new exact-head hosted Bazel CI #18380 pending; no current-head test result claimed. No local checkout/Bazel was available, so this is a reviewed patch awaiting execution, not a benchmark run. Representative same-corpus baseline/head one-file-edit wall time, actual source-read/hash work and peak RSS, independent human approval and applicable native proof tests remain unproven. Preserve merge hold; do not duplicate this issue PR, change protected decision-model/release lanes or weaken validation.
Published actual benchmark safety/reliability repair232d53af11afea1a30ed68b0e79dbeb11fdfb34d on existing PR #4846 (Refs #4507). Expected-old-head e4064bb, non-force fast-forward; independently verified PR head, branch ref, one-parent ahead-one/behind-zero compare and only two changed files (scripts/bench-hot-paths.py and its real-Git test).
Concrete defect: the one-file edit benchmark appended to the first git ls-files result without checking whether it was a symlink. A tracked symlink with an absolute target outside the disposable clone could therefore modify the operator's original source, despite the harness's isolation claim. Also, Git quotes paths containing whitespace; splitting by whitespace corrupted a valid tracked source name and aborted the measurement.
Fix: use NUL-delimited Git output, decode each pathname unchanged, and edit only a regular, non-symlink file whose resolved path remains inside the disposable clone. Preserve the existing pre-edit timer/RSS sampling, baseline generation fence, new-generation wait, and strict error handling.
Falsifiable regression: a real temporary Git repository commits an external-target symlink as the first candidate plus a spaced-name Rust file. The full fake-daemon benchmark must select the safe regular file, leave the external original untouched, preserve the original checkout, and clean its temporary run. On platforms without symlink fixture support, the spaced-path control still runs. Prior source would fail this test; it either errors on the quoted path or follows an unsafe candidate.
Validation boundary: hosted exact-head CI #18405 is pending. No passing new-head Python/Bazel/Rust tests or representative wall-time/peak-RSS measurements are claimed; no local checkout/Bazel was available. Independent approval and native tests when available remain required before merge. Issue/PR remain OPEN and unmerged. Separate #4520 and #4532 owners are unaffected; no Cargo, workflow change, force push, protected-model/release or production/infra action.
Published issue-owned benchmark readiness fix 7540e8c5f717 on existing PR #4846 (Refs #4507). The previous status_payload parsed only content[0].text, which can be a truncated/handle preview even for a successful tracedecay_status result. This caused the actual benchmark to miss the canonical ready generation and time out, blocking valid full-pass wall-time/RSS measurement. The shipped scripts/perf-gate.sh independently confirms the canonical structuredContent authority. The repair reads that typed document first and preserves the text-only envelope fallback.
Strengthened the existing real-Git/fake-daemon end-to-end benchmark regression: its successful status now deliberately returns complete freshness, generation, and memory only under structuredContent, with preview-only text. The previous parser cannot reach ready index/edit generations under this fixture; existing assertions require real nonempty generation transitions and memory rows. No test assertion, deadline, source integrity check, or workflow is weakened.
Verified GitHub branch/PR both at 7540e8c5f717291a7faebee1801a41b26e25093b, one-parent expected-head force=false update from 232d53af, compare ahead 1 / behind 0, only the two benchmark Python files changed. New exact-head hosted CI is queued; no new-head Python or Bazel test pass claimed. Local checkout/Bazel and representative original-corpus before/after time and peak RSS remain unavailable here, as do relevant native proof tests and independent human approval. Leave PR/issue open and merge hold in force; coordinate with PR maintenance without duplicate branches.
Activity
perf(code-graph): reread only re-sealed segments when proving a seal (#…
ScriptedAlchemy commentedon Nov 8, 2026
Issue #4507 ownership update (2026-11-07): the principal sealed-generation proof reuse fix was already merged via #4523. I opened draft #4791 against
masteratfc4322c74abfd7d2beee88c8f70278c45deef123to add a falsifiable real one-file-successor proof regression (cold versus parent-warm check counts, plus existing corruption refusal). This is test coverage, not a second copy of the production fix. Separately, #4520's sparse-successor CPU/RSS residual has active draft #4789; I have not modified that overlapping lane. Hosted Bazel CI and independent review are still pending; no native macOS runner/local Bazel is available in this run. Keep #4507 open until full end-to-end time/memory and corruption acceptance is demonstrated onmaster.ScriptedAlchemy commentedon Nov 8, 2026
Incremental #4507 engineering update (2026-11-07): PR #4791 now has commit
e16aaf0. In addition to cold-vs-parent-warm nonzero inherited-segment proof checks on a real one-file successor, it now warm-proves the successor, mutates its newly changed sealed file segment without changing length, requires typedCorrupt, restores the bytes and verifies recovery. This closes the blind spot where a verifier could have passed the inherited-read count assertion while skipping all changed source. PR moved to ready-for-review solely to trigger standard Bazel Linux CI; the general PR sweep has an explicit do not merge hold on #4791 pending exact-head CI and independent review. CI: https://github.com/ScriptedAlchemy/tracedecay/actions/runs/42434575039 (queued at this check); native macOS not part of the default run. No releases, Kev/decision-model merges, or overlapping #4520 source edits. Issue remains open pending end-to-end memory/time acceptance onmaster.fix(code-graph): retain authenticated proofs across interrupted reads
ScriptedAlchemy commentedon Nov 8, 2026
Published a further production fix in the existing #4791, commit
cd1e40aada9d38801877ea62b92520d1daaf4f0a: preserve the last authenticated sealed-source snapshot across interrupted and overlapping proofs, without exposing tentative hashes or allowing stale completion to replace a newer successful proof. Four real-file regressions cover that reuse and the corruption/root/extent/deletion boundaries; the existing real one-file successor proof regression is retained.No duplicate PR or overlapping #4520 implementation was created. Branch update was an expected-head, non-force fast-forward from
216e726ef58b46ce6bcc4017e1ac225a51fcf8c1.Exact-head hosted Bazel CI is queued/pending at https://github.com/ScriptedAlchemy/tracedecay/actions/runs/42442313648; native macOS is skipped. No local Bazel is available, and no Cargo was run. Nonzero passing regression results, independent review and fresh full-pass time/RSS evidence are not yet established. #4791 remains under an explicit merge hold; this issue remains open. Other sweeps should coordinate on that PR rather than duplicating the fix or treating partial evidence as acceptance.
ScriptedAlchemy commentedon Nov 8, 2026
#4507 exact-head verification evidence (2026-11-08 UTC): Downloaded and inspected the actual hosted Bazel test artifact for existing PR #4791 at
cd1e40aada9d38801877ea62b92520d1daaf4f0a.//crates/tracedecay-store-runtime:unit_testpassed 177 tests, 0 failed, 2 ignored, 0 filtered, including all four newly added real-file verification tests andinherited_segment_proof_rejects_wrong_manifest_extentfor the real one-file successor. This is a non-vacuous execution result, not just a green build. Test artifact: https://github.com/ScriptedAlchemy/tracedecay/actions/runs/42442313648/artifacts/13606073424.New blocker: Overall CI #17306 is red despite successful repository formatting, Clippy, feature gates, dashboard, and the store-runtime tests. The three failing Linux test targets are automation-runtime (2 timeout assertions), runtime acceptance (LSP deadline), and automation-runner (3 scheduler/backend assertions). The latter three individual automation-runner failures are also present on unmodified
master(CI #42328697300). The other failures are outside this PR's changed paths, but require separate baseline/flakiness attribution—no test weakening or bypass. Windows was still running, native macOS skipped, and independent review plus fresh real-workload full-pass time/peak-RSS measurements remain missing.Updated the existing #4791 description with this evidence and kept its explicit merge hold. No second PR, code rewrite, Cargo execution, release, Kev/decision-model merge, or issue closure.
Refs #4507remains appropriate; not yetFixes.test(code-graph): prove warm recovery after corrupt successor
ScriptedAlchemy commentedon Nov 8, 2026
#4507 production-regression follow-up published to existing PR #4791:
cb58c6d. Safely fast-forwarded from exact old headcd1e40awithforce=false; no duplicate PR or unrelated source change.New falsifiable real-path coverage: The prior one-file successor test checked that corruption of a changed sealed segment is refused and restoration succeeds, but did not prove that failed verification preserves authenticated unchanged source. It now counts production proof callbacks after the corrupted changed segment is restored: recovery must do a nonzero new proof, and on Unix a cold fresh verifier must perform at least one extra proof check per inherited file segment. This fails if the last good proof is evicted on error, so source integrity and incremental recovery cost are asserted together. Non-Unix still requires full proof.
Previous exact-head (
cd1e40a) hosted Bazel artifact had 177 passing store-runtime tests including 5 #4507 regressions; that is not validation of newcb58c6d. New head CI #17387 is pending. Earlier overall Linux CI #17306 had unrelated automation/runtime acceptance failures; do not weaken or bypass them. Native macOS validation, independent review and fresh full-pass wall/peak-RSS measurements are still missing.The merge hold remains, this issue remains OPEN with
Refs #4507(notFixes), and the separate #4520/#4789, protected draft #4776 and read-only release #4763 are untouched. No Cargo, force update, sweep merge, release, or infrastructure action.ScriptedAlchemy commentedon Nov 8, 2026
New exact-head CI blocker for #4507 (2026-11-08 09:04 UTC): PR #4791 run #17387 at
cb58c6dd7e0806637eb63388ed229c67eb98046apassed repository gates, change-scope and benchmark-harness checks, but Clippy and Dashboard failed during Bazel dependency analysis because GitHub returned HTTP 504 fetching the pinnedhermeticbuild/hermetic-launcherrunfiles-stub-x86_64-linuxandfinalize-stub-x86_64-linuxassets. No compiler/test failure was reached for those jobs. Linux Bazel and feature gates remained queued. Native macOS was skipped.I attempted to rerun the Clippy job directly, but GitHub refused with HTTP 403 because the containing workflow is still running. The retry must happen only when the run finishes; no workflow/infrastructure alterations or speculative source commits are warranted for this external 504. The PR description is updated with exact log-derived attribution. Its MERGE HOLD and
Refs #4507remain, with independent review, nonzero exact-head regression results and fresh full-pass wall/peak-RSS measurements still outstanding. Do not duplicate this issue PR, bypass checks or close #4507.ScriptedAlchemy commentedon Nov 8, 2026
Additional exact-head #4507 CI blocker (2026-11-08 09:40 UTC): #4791 run #17387, head
cb58c6dd7e0806637eb63388ed229c67eb98046a, also lost its Linux fleet Bazel build/test lane before tests could execute. Linux fleet job 128671373185 reports the same pinnedhermeticbuild/hermetic-launcherGitHub release-asset HTTP 504 Gateway Time-out, now for bothfinalize-stub-x86_64-linuxandrunfiles-stub-x86_64-linux. Itsbazel build --config=ci //...ended in Bazel repository-loading errors; no exact-head store-runtime test result is available. The dependent Linux gate failed because its fleet lane failed; this is not an additional source-test failure. Repository gates, benchmark-harness checks, change-scope and feature gates passed; Clippy/Dashboard previously failed the same upstream 504. The Windows job was still running at the latest poll, so the workflow remains in progress and the failed-job retry is not yet eligible. Native macOS was skipped.No speculative source or infrastructure adjustment is justified by a third-party fetch outage. Keep
Refs #4507, the #4791 merge hold, and #4507 open. Retry failed jobs through GitHub's native workflow retry only after the run completes; retain nonzero regression, independent review and fresh wall-time/peak-RSS evidence requirements.test(code-graph): prove restored source is actually rehashed
ScriptedAlchemy commentedon Nov 8, 2026
Published a falsifiability fix to the existing #4507 PR #4791:
25caf8d. The previous post-corruption assertionrecovered_reads > 0could pass on the manifest/lifecyclecheckcallbacks even if no repaired source segment was rehashed. The actual test now records a fully warmed successful baseline, demands a strictly higher callback count when verifying the changed segment after same-length corruption and restoration, and asserts a subsequent successful proof returns to the warm baseline. Existing cold-vs-inherited-digest bounds, typed corruption refusal, and cross-platform fallback stay intact. One file changed, tests only, no weakening or production/infrastructure change.This was a verified non-force fast-forward of
test/4507-inherited-segment-prooffrom exact expected SHAcb58c6dd7e0806637eb63388ed229c67eb98046a; no competing #4507 PR was created. New exact-head hosted Bazel CI #17439 is pending, not a claimed pass. Previous head #17387 lost Linux/Clippy/Dashboard to pinned hermetic-launcher GitHub HTTP 504 fetch errors. Previouscd1e40astore-runtime executed 177 passing tests but does not validate25caf8d. Prior Windows job for #17387 was still building, preventing native failed-job retry then; hosted macOS remains skipped.MERGE HOLD: No independent review or real 6.6k-file full-pass wall-time/peak-RSS evidence yet. Keep #4507 open,
Refs #4507only. #4520/#4789 are separate, draft Kev/decision-model #4776 is protected, release #4763 read-only. No Cargo, force update, auto-merge, Kody, Copilot or external delegation.test(code-graph): honor non-Unix sealed proof fallback
ScriptedAlchemy commentedon Nov 8, 2026
#4507: published a cross-platform regression correction to existing PR #4791 (2026-11-08 UTC).
Commit
1e530bffixes a newly identified contradiction in the real one-file successor test. The previous assertion requiredrecovered_reads > warmed_readson every platform. ButSealedSegmentFingerprintV1::ofdeliberately returnsNoneon non-Unix systems, so warm and recovered proofs both rehash all segments and must produce equal callback counts. The test now retains the strict positive recovery-vs-warm proof delta on Unix (where caching exists) and asserts equal full-read costs on unsupported platforms. Typed corrupted-segment refusal, cold-vs-inherited nonzero bounds, and successful rewarming assertions remain unchanged. No production behavior, weakening of Unix acceptance, CI/infrastructure, or dependency changes.Pushed an exact-parent, non-force fast-forward from
25caf8d700236e7bcf0ee07515f3a22bf5a7ade1to1e530bf2aedabbc8caaf5dcca5d24a85eb015ccfon the existingtest/4507-inherited-segment-proofbranch; verified the new remote ref and one-file commit. Exact-head hosted Bazel CI #17446 is queued; no test pass is claimed for this head. Local Bazel/Bazelisk is unavailable; Cargo was not run.MERGE HOLD stays. Need nonzero exact-head regression execution, independent review, and fresh real-workload wall-time/peak-RSS; native macOS remains skipped by the default workflow. Keep
Refs #4507, issue open, no duplicate PR. #4520/#4789, protected draft #4776, and release #4763 remain outside this change.30 remaining items
fix(bench): include source mutation in one-file edit measurements
ScriptedAlchemy commentedon Nov 10, 2026
Published issue-owned benchmark measurement fix
a6bbbbbb6d47on existing PR #4846 (Refs #4507). The existingscripts/bench-hot-paths.pyalready exercises a real disposable Git clone and daemon through an appended one-file edit,sync, and a new ready generation, with wall time and daemon peak RSS. But it previously started its wall timer and /proc RSS sampler after writing and closing the edited source file. Filesystem-watch-triggered refresh can therefore start before observation, systematically undercounting the very incremental work #4507 needs measured.The patch moves timer and sampler activation ahead of the source mutation and keeps stop after the observed new generation is ready; sampling is stopped in
finallyon error. Added nonvacuoustest_edit_timer_and_rss_observer_cover_the_file_mutationthat drives the actual harness edit method against a tracked real file with a deliberately slow append, asserts sampling begins before the edit and stops after generation readiness, requires measured wall time to include the delay, and checks a positive peak-RSS result. Existing full fake-daemon end-to-end harness tests are preserved. No production graph/cache source change, deadline increase, weakened assertion or extra test-only production path.Published safely via exact expected old head
85c8cd0, nonforce; independently verified PR + named ref ata6bbbbbb6d4740ea7949ff80b8888918ca973ccf, compare ahead 1 / behind 0, exactly two script/test files. Exact-head CI #18355 is queued, not passed. This is an accuracy fix for obtaining future same-corpus before/after evidence, not itself a representative performance measurement or proof that source rereads are eliminated. Rust Bazel tests/native validation, original ~6.1k-file same-host wall/peak-RSS comparison, and independent approving review remain outstanding. Keep #4846 unmerged and #4507 open; do not duplicate this work in sweeps. No Cargo, external delegation, protected model/release changes, production or infrastructure work.fix(bench): fence one-file measurements to the pre-edit generation
ScriptedAlchemy commentedon Nov 10, 2026
Published e4064bbcb350 on existing #4846 (
Refs #4507), expected-head non-force fast-forward froma6bbbbbb6d47. Independent PR and branch readbacks agree on the new SHA; compare confirms exactly one commit ahead / zero behind and two files,scripts/bench-hot-paths.pyplusscripts/test-bench-hot-paths.py.Actual measurement-correctness defect: the one-file benchmark waits for a generation different from the ID saved at initial indexing, but the daemon can complete another background refresh before the measured edit. The pre-existing generation ID then becomes stale; the old implementation could accept that unrelated already-completed refresh as the edit's result, reporting an artificially short pass and missing incremental peak RSS. A response claiming current/ready with an absent generation ID could also satisfy the old
None != old_idcheck.Repair: require a nonempty, current/ready generation ID immediately before mutating the tracked source, use that freshly observed ID as the baseline for the edit-to-ready wait, require a nonempty successor ID, and update the tracked generation after the measured pass. Existing timer/RSS observer still starts before the file mutation and stops after readiness, preserving the previous fix.
Falsifiable tests: the existing real tracked-file slow-edit regression now supplies an obsolete initial generation and an independently newer pre-edit generation; it asserts the wait uses the newer ID, includes the deliberate edit latency, and records peak RSS. It also requires a missing pre-edit generation to refuse without changing the source or starting sync, and a missing post-edit generation to be skipped until a later valid successor arrives (two status polls). Previous full fake-daemon harness tests remain unchanged.
Validation: new exact-head hosted Bazel CI #18380 pending; no current-head test result claimed. No local checkout/Bazel was available, so this is a reviewed patch awaiting execution, not a benchmark run. Representative same-corpus baseline/head one-file-edit wall time, actual source-read/hash work and peak RSS, independent human approval and applicable native proof tests remain unproven. Preserve merge hold; do not duplicate this issue PR, change protected decision-model/release lanes or weaken validation.
fix(benchmark): fence edits to regular clone-owned files
ScriptedAlchemy commentedon Nov 10, 2026
Published actual benchmark safety/reliability repair
232d53af11afea1a30ed68b0e79dbeb11fdfb34don existing PR #4846 (Refs #4507). Expected-old-heade4064bb, non-force fast-forward; independently verified PR head, branch ref, one-parent ahead-one/behind-zero compare and only two changed files (scripts/bench-hot-paths.pyand its real-Git test).Concrete defect: the one-file edit benchmark appended to the first
git ls-filesresult without checking whether it was a symlink. A tracked symlink with an absolute target outside the disposable clone could therefore modify the operator's original source, despite the harness's isolation claim. Also, Git quotes paths containing whitespace; splitting by whitespace corrupted a valid tracked source name and aborted the measurement.Fix: use NUL-delimited Git output, decode each pathname unchanged, and edit only a regular, non-symlink file whose resolved path remains inside the disposable clone. Preserve the existing pre-edit timer/RSS sampling, baseline generation fence, new-generation wait, and strict error handling.
Falsifiable regression: a real temporary Git repository commits an external-target symlink as the first candidate plus a spaced-name Rust file. The full fake-daemon benchmark must select the safe regular file, leave the external original untouched, preserve the original checkout, and clean its temporary run. On platforms without symlink fixture support, the spaced-path control still runs. Prior source would fail this test; it either errors on the quoted path or follows an unsafe candidate.
Validation boundary: hosted exact-head CI #18405 is pending. No passing new-head Python/Bazel/Rust tests or representative wall-time/peak-RSS measurements are claimed; no local checkout/Bazel was available. Independent approval and native tests when available remain required before merge. Issue/PR remain OPEN and unmerged. Separate #4520 and #4532 owners are unaffected; no Cargo, workflow change, force push, protected-model/release or production/infra action.
fix(benchmark): parse canonical structured status for readiness
ScriptedAlchemy commentedon Nov 10, 2026
Published issue-owned benchmark readiness fix
7540e8c5f717on existing PR #4846 (Refs #4507). The previousstatus_payloadparsed onlycontent[0].text, which can be a truncated/handle preview even for a successfultracedecay_statusresult. This caused the actual benchmark to miss the canonical ready generation and time out, blocking valid full-pass wall-time/RSS measurement. The shippedscripts/perf-gate.shindependently confirms the canonicalstructuredContentauthority. The repair reads that typed document first and preserves the text-only envelope fallback.Strengthened the existing real-Git/fake-daemon end-to-end benchmark regression: its successful status now deliberately returns complete freshness, generation, and memory only under
structuredContent, with preview-only text. The previous parser cannot reach ready index/edit generations under this fixture; existing assertions require real nonempty generation transitions and memory rows. No test assertion, deadline, source integrity check, or workflow is weakened.Verified GitHub branch/PR both at
7540e8c5f717291a7faebee1801a41b26e25093b, one-parent expected-headforce=falseupdate from232d53af, compare ahead 1 / behind 0, only the two benchmark Python files changed. New exact-head hosted CI is queued; no new-head Python or Bazel test pass claimed. Local checkout/Bazel and representative original-corpus before/after time and peak RSS remain unavailable here, as do relevant native proof tests and independent human approval. Leave PR/issue open and merge hold in force; coordinate with PR maintenance without duplicate branches.fix(benchmark): prove semantic one-file edits reached ready graph
fix(benchmark): fence one-file latency to the served graph
perf(code-index): expose sealed source proof work per edit
fix(bench): sample macOS daemon RSS during incremental edits
fix(bench): terminate orphaned daemon process groups
fix(bench): preserve Python source file modes
fix(bench): keep harness test within its Bazel size