Release 0.41.0: one producer per language (#309 flip of six languages), declared profiles honoured, silent-watcher detection, removed worktrees stay removed (#361 #310), core release bundles the published channel (#318 #354) #389

Merged
buildagent merged 29 commits from train-0410 into master 2026-10-16 19:55:11 +02:00
Member

Summary

One producer per language (#309). csharp, javascript, php, python, rust and typescript are package-owned under auto (PACKAGE_OWNS_UNDER_AUTO). Ruby stays on the builtin until its declared deltas are attributed.

The flip is proven by identity:

  • Every corpus suite indexes through the shipped first-party packages.
  • baseline.json and cost-baseline.json reproduce byte for byte.
  • precision_gate is 7/7 with phantom_count 0 and unchanged POPULATION.
  • The stage/tier3 records changed only by producer-key renames with identical counts (operator-approved).
  • The host-extraction-epoch record follows the pins, which moved to the shipped versions (operator-approved).

Lanes

Each lane was verified by the integrator, with mutations re-run.

  • O, oracle. The corpus gates and precision_gate grade the shipped producers. The producer list is the product's own (PACKAGE_OWNS_UNDER_AUTO), never a second table.
  • A, adoption.
    • A start adopts the bundled first-party packages for flipped languages. Only a package signed by a key the binary ships qualifies.
    • A third-party package, or one anchored with --first-party, is never adopted.
    • A build with no builtin and no healthy package says the language is not indexed.
    • One conform::admit replaces four copies of install → C1 → verdict.
  • P, profiles.
    • Four readers that bypassed declared profiles now read them.
    • Five per-language tables became declarable profile traits.
    • One receiverless-refusal helper.
    • Corpus-neutral, and cost below the blessed values.
  • K. A first-party package's declaration governs the rows it owns, and only while it owns them.
  • W, watcher (#381 follow-up, field report from zammad-dev).
    • A canary liveness probe; silent native delivery falls back to notify's PollWatcher, and the fallback is disclosed.
    • Served paths are stat-checked and re-queued.
    • touch re-decides a row that disagrees with disk.
  • X (#361, #310). A removed worktree is dropped and never recreated (the cause was the daemon's exit log line recreating .code-index/). A new worktree is linked mid-session.
  • R, release/CI (#318, #354).
    • The core release bundles exactly what the signed pkg-channel published.
    • The version bump never edits a package row: this bump touched 2 catalog lines, against 20 for v0.40.
    • Green is inherited by tree on merges.
    • The Ruby cost record lives beside its suite.

Integration fixes

  • The flip grew project_overview. It was trimmed, not raised: an absent bundle is not reported per start, and not-adopted languages are grouped by reason.
  • The overview's watcher field is bounded.
  • Gate registries were updated for the new constants and tests.
  • The release-only seam test now runs in the CI corpus job.

Local verification (one full run, including the nightly-only jobs)

  • fmt, clippy (linux and windows-gnu), and docs: all green.
  • nextest workspace: 5186/5186. Daemon leg: 979/979. ci-timing: 37/37.
  • All corpus suites, tier 3, precision_gate, all 11 package full legs, and the 7 overhead legs: green.
  • Corpus scale, mixed load, overview 2m, the packages-only build plus phase-2 acceptance, plugin-path cost, and every bench: green.
  • No package bytes changed, so the core release needs no pkg/ tag first.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj

## Summary **One producer per language (#309).** csharp, javascript, php, python, rust and typescript are package-owned under `auto` (`PACKAGE_OWNS_UNDER_AUTO`). Ruby stays on the builtin until its declared deltas are attributed. The flip is proven by identity: - Every corpus suite indexes through the shipped first-party packages. - `baseline.json` and `cost-baseline.json` reproduce byte for byte. - `precision_gate` is 7/7 with phantom_count 0 and unchanged POPULATION. - The stage/tier3 records changed only by producer-key renames with identical counts (operator-approved). - The host-extraction-epoch record follows the pins, which moved to the shipped versions (operator-approved). ## Lanes Each lane was verified by the integrator, with mutations re-run. - **O, oracle.** The corpus gates and `precision_gate` grade the shipped producers. The producer list is the product's own (`PACKAGE_OWNS_UNDER_AUTO`), never a second table. - **A, adoption.** - A start adopts the bundled first-party packages for flipped languages. Only a package signed by a key the binary ships qualifies. - A third-party package, or one anchored with `--first-party`, is never adopted. - A build with no builtin and no healthy package says the language is not indexed. - One `conform::admit` replaces four copies of install → C1 → verdict. - **P, profiles.** - Four readers that bypassed declared profiles now read them. - Five per-language tables became declarable profile traits. - One receiverless-refusal helper. - Corpus-neutral, and cost below the blessed values. - **K.** A first-party package's declaration governs the rows it owns, and only while it owns them. - **W, watcher (#381 follow-up, field report from zammad-dev).** - A canary liveness probe; silent native delivery falls back to notify's `PollWatcher`, and the fallback is disclosed. - Served paths are stat-checked and re-queued. - `touch` re-decides a row that disagrees with disk. - **X (#361, #310).** A removed worktree is dropped and never recreated (the cause was the daemon's exit log line recreating `.code-index/`). A new worktree is linked mid-session. - **R, release/CI (#318, #354).** - The core release bundles exactly what the signed pkg-channel published. - The version bump never edits a package row: this bump touched 2 catalog lines, against 20 for v0.40. - Green is inherited by tree on merges. - The Ruby cost record lives beside its suite. ## Integration fixes - The flip grew `project_overview`. It was trimmed, not raised: an absent bundle is not reported per start, and not-adopted languages are grouped by reason. - The overview's watcher field is bounded. - Gate registries were updated for the new constants and tests. - The release-only seam test now runs in the CI corpus job. ## Local verification (one full run, including the nightly-only jobs) - fmt, clippy (linux and windows-gnu), and docs: all green. - nextest workspace: 5186/5186. Daemon leg: 979/979. ci-timing: 37/37. - All corpus suites, tier 3, `precision_gate`, all 11 package full legs, and the 7 overhead legs: green. - Corpus scale, mixed load, overview 2m, the packages-only build plus phase-2 acceptance, plugin-path cost, and every bench: green. - No package bytes changed, so the core release needs no pkg/ tag first. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
CI grades a tree, not the commit object around it. The merge commit of an
up-to-date PR has its head's tree byte for byte, yet runs 1174-1176 re-ran on
master a48fb945 everything 1171-1173 had just passed on 960f4704.

require_ci_green.sh: a job with no row of its own (absent / skipped only) is
decided by the newest non-skipped row of any other commit in the snapshot
whose tree (read locally: Forgejo's /git/commits answers tree.sha with the
commit sha) equals this one's. Own rows still decide; a newer red on the
same tree wins; a tree one byte different has no twin. The release gate and
ci_changes.sh's already-green decision (now scoped to WORKFLOW_ID) both get
it, so ci.yml and ci-windows.yml need no YAML change. package-ci.yml's
changes job now passes GREEN_JOBS (every leg, expanded) and the API env.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tests/corpus/ruby-package-cost.json -> crates/first-party-tests/tests/ruby_package_cost.json,
byte-identical (100% rename; no number moved). Its only reader is
ruby_package_cost.rs, which only `package (ruby)` runs; under tests/corpus/
a re-bless was a host path and ran the whole host CI (the #384 schema
re-measure did). Beside the suite, ci_changes.sh routes it to the ruby leg
and no host group. bless_registry follows: the registered artifact and the
vm_step census. The bless template's crate references are corrected for
the next bless (the committed record keeps its old _comment until then).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
version_bump.sh: every package asset URL (`/releases/download/…/<file>.cip[s]`,
and the own-release `pkg%2F…` spelling it subsumes) is not a version site —
one pattern (PKG_URL_ERE) for the scan, the rewrite and --version-only. The
v0.40 bump rewrote twenty such lines; since a core release no longer packs
packages, a rewritten row would name a release that does not carry it.

release_registry.sh: the retag is deleted. A row's URL is copied as the
bootstrap snapshot states it, and the channel decides membership for EVERY
row (published_rows), not only own-release ones; values still come from the
source, so a lying channel can only make a release list less. The audit's
URL rule is "a release download ending in its asset"; downloadability is
asked of the host before publication.

registry_schema_gate: a row's segment is its own release or a core release
no newer than code_index.tag.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
release.yml loses its `package-plugin` matrix (pack, sign, install path and
negative controls for all eleven packages on every core release). A package
is released only by its own pkg/<id>/v<version> tag through
package-release.yml, which already ran the same scripts.

assemble-archives now fetches the published package channel and admits the
seven bundled language packages from it with the shipped linux binary
(release_bundle.sh fetch): the channel is verified through the binary's own
catalog door, each listed URL must download anonymously
(package_channel_check.sh --downloadable), and each .cip must pass `plugin
install --sha256 <the channel's digest>` (verifies the .cips against the
compiled-in anchors; executes nothing). BUNDLED_PACKAGES is a literal held
equal to every tests/packages/<builtin-language> by release_gate.

The release job uploads no package asset, its audit no longer requires
.cip/.cips/.pub assets, it fetches EVERY catalog row before publishing
(--downloadable-new <catalog> -; the --downloadable-own mode is deleted), and
the notes replace eleven per-package install blocks with the bundle (read out
of the archive) and the owner-switch commands.

Tests: release_gate / release_fixture_gate retarget their packing,
signing, install-path and negative-control checks to package-release.yml's
`package` job; new the_core_release_packs_no_package_and_bundles_what_the_
channel_published and the_bundle_is_every_language_package; the bundle fetch
is executed against a fake host in package_release_scripts. tests/packages/
README.md carries the "fix one package -> release it" runbook.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The five per-language tables become ProfileTraits fields, declarable in
`[languages.profile]` and defaulting to the builtin through `extends`:
QUALIFIED_NAME_SEPARATORS -> qualified_name_separator (NameSeparator),
IMPORTLESS_PROFILES -> importless, RECEIVER_LOCALITY_BLIND_PROFILES ->
receiver_locality_blind, IDENTIFIER_SUFFIXES -> identifier_suffixes
(IdentifierSuffixes), kinds::POSITION_TYPE_KINDS_BY_LANG ->
position_type_kinds (PositionTypeKinds). The tables are deleted; each
rationale moved to its field. The resolver reads the three SQL answers
through temp.lang_profile (new columns importless,
receiver_locality_blind, position_type_module, and the separator as a
TEXT column), so no language id is spelled in SQL any more.
cs_scope_fqn joins with the profile's separator and leaves out a
partial-types language whose names do not compose.

Readers that answered from the compiled builtin now read the index's
declarations: local_index's two receiverless name-fallback readers
(book_or_builtin), and mcp-server's member attribution and
check_rename suffix check, through a new `declared_profiles` index
method (IndexAccess default None, LocalIndex impl, RPC arm + client,
rpc_coverage leg) and `code_index_daemon::profile_book`. The writer and
the packaged-qualifier join take the PRODUCER's declaration
(ParseOutcome::profile, FactContext::profile), because the index's
declared_profiles table is written only after a pass extracts.

The three #378 refusal DELETEs go through one helper,
refuse_receiverless_decisions(tx, table, method_pool), same SQL.

first_party_profiles' canonical declarations and the manifest's
restated vocabulary grow to 24 fields; every builtin answer is
transcribed and graded in the_builtin_answers_are_the_inline_tests_they_replaced.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
- bundled::adopt: a daemon / --no-daemon MCP start adopts the bundled package
  of each enabled `auto` language for which `auto` means the package in this
  build (languages::Build::auto_means_package: flipped, or no builtin
  compiled). It runs the plugin-add door's statements (install, C1, passing
  verdict, requested grant) with no question, licensed ONLY by a key this
  build ships (Store::is_first_party_key, key bytes). An operator-vouched
  `--first-party` key is offered, never adopted. Withdrawn (plugin disable)
  and already-served languages are left alone. Disclosed once on
  project_overview.language_ownership.adoption; nothing due costs nothing.
- No builtin, no fallback: owner_rows lists every enabled language the build
  has no builtin for; nothing producing it is owner `none` with reason + fix.
  Said by project_overview (own semantics), `languages list`, `doctor`, and
  the server instructions (package-only languages line).
- `[languages] <lang> = "builtin"` on a build without that builtin is
  refused at parse (so every reader and writer refuses it).
- builtin_compiled reads the compiled plugin set, per row language.
- tests: bundled_adopt.rs (first-party adopted under a test-local flip list,
  third-party never adopted, withdrawn not re-adopted), languages unit tests,
  MCP renderer test, phase2_acceptance no-fallback check on the infra-only
  binaries (real-binary third-party refusal).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
conform::admit owns the tail every door runs after its own consent:
install the consented bytes, run C1 under the grant about to be made,
require the RECORDED verdict to cover it. It returns the stage it stopped
at with what it had done (installed package, verdict), so each door keeps
its own wording. Callers: `plugin add` (cmd_add), `plugin update`'s apply
(a fourth copy found while doing this), the MCP `plugin_add` tool, and
bundled::adopt_one. CLI output and MCP payloads are unchanged; the CLI
prints through one admit_printing/print_verdict pair shared with
`plugin check`.

The heads stay per door (where the bytes come from, the pin and its
message, the human confirmation, which needs the grant and the record set
BEFORE consent to measure the reindex domain); the adoption caller keeps
the shipped-key check, asked of the bytes it hands to admit.

Tests: bundled_adopt gains a third party's displacing package already
approved (not adopted over) and a first-party package failing C1 (not
adopted).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Joining temp.lang_profile into the scope skeleton cost cs-dapper +25.5%
vm_step (cost_attribution: 6.36M for this one statement; 33.9M with the
order forced). The separator is now a PK lookup in the recursive step and
the NULL-separator exclusion rides in sk's existing IN-subquery; the pass
measured 23.38M against the blessed 23.53M.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
One table, `code_index_test_support::shipped::PRODUCERS`, names the
producer of every language a gate grades: the pinned first-party package
for csharp, javascript, php, python, rust and typescript; the builtin for
Ruby until its deltas are attributed. `shipped::install` is the one way a
test installs those packages (frozen PINS bytes, production install and
approval doors, fixture key classified first party); package_overhead
now delegates to it.

- Corpus harness: `corpus::index_repo` / `corpus::shipped` index every
  pinned repo through the table (builtin dropped for package-produced
  languages, packages served by a real PackageHost). Every corpus suite
  (ratchet, stage, cost, metamorphic, mutation, watcher replay, scale,
  tier3, cost_attribution) uses it. corpus_stage and corpus_tier3
  check, before any count, that the active generation was built with
  exactly the pinned packages (replaces the builtins-only digest gate).
- precision_gate: each daemon gets its fixture language's pinned package
  installed and owning the language (`[languages] <lang> = "package"`),
  and the gate fails unless every language was produced by the table's
  producer.
- PINS: the six languages move to the bytes the registry ships
  (csharp/javascript/php/python/typescript 0.1.3, rust 0.1.4). The old
  0.1.1 rust snapshot is NOT identical to the builtin (-140 resolved on
  rust-ripgrep); the shipped bytes are.
- ci-windows.yml: precision_gate runs once more with its output, so the
  Windows log shows the producer and population it graded.

Identity: baseline.json and cost-baseline.json reproduce unchanged;
precision_gate 7/7, phantoms 0, POPULATION lines identical.
NOT re-recorded (operator decision): stage-baseline.json and
tier3-baseline.json move only by `producer.builtin/<lang>` ->
`producer.de.h-dv.<lang>/<lang>` at identical counts;
host_extraction_epoch.json moves by the pinned package digests (rust's
fingerprint with its new fixture).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Field report (zammad-dev): the native watcher was armed and delivered
nothing; a deleted file stayed served because only NEGATIVE replies ever
compared the tree against the rows. In-house symptom (lane A): a row the
writer invalidated stayed stale with no event left to re-decide it, and
`touch` did not help.

- Liveness probe (indexer watcher): at arm-up and each periodic reconcile
  a canary written into the state directory must come back through the
  native backend within PROBE_BOUND + debounce. No event: warn once,
  re-walk, and follow the tree with notify's PollWatcher (one
  non-recursive poll per watchable directory, interval scaled with the
  tree: 1 ms/dir within 2-30 s) until a later probe succeeds. A state
  directory outside the tree (--db) or on another filesystem is
  `unprobed`, and says so.
- Disclosure: Stats/project_overview `watcher`, doctor `watcher delivery`,
  daemon.log WARN on the transition, one line in server instructions.
- A served path is checked: every reply's served paths (`path`/`file`
  rows, bounded) are stat-checked by their index against their rows via
  the new `served_paths` RPC; a path gone or changed is re-queued through
  the existing recheck queue and marked (`lagging`,
  `served_paths_lagging`).
- `Modify(Metadata)` events are re-decided when the stat disagrees with
  the row (was skipped unconditionally): `touch` heals, and the poll
  backend's mtime events land.
- Test seam CODE_INDEX_TEST_DROP_NATIVE_EVENTS is read only under
  cfg(debug_assertions); a release-only test proves it inert.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
stage-baseline and tier3-baseline: only producer.builtin/<lang> keys renamed to
producer.de.h-dv.<lang>/<lang>, counts identical (cs-dapper 157, js-express 141,
php-guzzle 131, python-flask 83, rust-ripgrep 110, ts-zod 4+401; py-django and
rust-analyzer in tier3). No rule./stage./influence./count. key moved.
host_extraction_epoch: PINS move from the stale frozen rust 0.1.1 to the shipped
0.1.3 (x5) and rust 0.1.4; five fingerprints unchanged, rust 11->12 fixtures.
baseline.json and cost-baseline.json are byte-identical.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj

# Conflicts:
#	crates/indexer/tests/corpus_watcher_replay.rs
A first-party package stamps its rows with the reserved plain id
(`de.h-dv.rust/rust` -> `rust`), but its `[languages.profile]` was stored
in `declared_profiles` under the wire id, and every reader (the resolver's
`temp.lang_profile`, `local_scope`, `book_or_builtin`, the daemon's
`declared_profiles` / `profile_book`) asks by a row's lang. So the
declaration never reached the rows it was written for; the builtin
answer stood.

One rule, `profiles::governed_lang`, applied where the live set's
declarations are materialised (`profiles::governing`, called by
`write_capability_grants`), so every reader inherits it: a declaration
is keyed by the language its producer's rows carry -- the wire id for a
third-party language, the plain id for a first-party one while that
package OWNS the language (it displaces the builtin after the project's
owner setting narrowed `displaces`, or this build has no builtin for it).
When the builtin owns the language the package's declaration governs no
row and is not stored.

Corpus-neutral: first-party declarations equal the builtin
(first_party_profiles). corpus_ratchet, corpus_stage, corpus_cost green;
precision_gate 7/7, phantoms 0; no file under tests/corpus/ moved.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
The linked-project set is re-derived from disk by the startup discovery
(derive_links) when a reply needs it, on a 2 s bound (project_set). A
link still found keeps its handle; a new one is attached by the startup
task (attach_link); one no longer found is dropped from fan-out,
answer_provenance and linked_projects and answers "removed" instead of a
git `fatal:`.

Nothing creates a root that does not exist: the state directory
(state_fs::ensure_state_dir) creates only itself and refuses a missing
parent; the daemon log append goes through the same rule instead of
create_dir_all (the write that recreated removed worktrees on exit); the
one spawn gate (lifecycle::spawn_refusal) refuses a missing root; and
daemon rediscovery stops for a root that is gone.

Tests: two e2e (real git worktrees in a tempdir; daemon leg for the
removal), unit tests for each refusal and the merge rule; every mutation
run red, restored by md5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PACKAGE_OWNS_UNDER_AUTO lists the six languages proven identical through
their shipped packages (lane O: every corpus record reproduced). Ruby stays
on the builtin until its declared deltas are attributed.

The flip grew project_overview past its content ceiling (+85 tokens): every
start of a binary with no bundle reported each flipped language as
not_adopted with the bundle path. Trimmed, not raised:
- a binary with NO code-index-packages/ beside it (a source build) has
  nothing due: the builtin keeps the language as before the flip, and
  nothing is reported per start; a bundle that is present and lacks or
  refuses a package is still reported (an_absent_bundle_is_silent_and_an_
  incomplete_one_is_said; mutation: drop the dir_present return -> RED);
- not_adopted languages with the same reason and witness are said once
  (languages_refused_alike_are_said_once; mutation: one row per entry -> RED);
- the no-bundled-package witness is empty (its reason code says it).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
- code_index_test_support::shipped derives producer_of/package_languages from
  code_index_core::lang_profile::PACKAGE_OWNS_UNDER_AUTO instead of a second,
  literal table (lang_literal_gate was red on it, and two copies of the flip
  decision can drift). Mutation (RUN): package_languages gains kotlin -> RED.
- host_reads_no_live_package: release_gate 30->27, release_fixture_gate 1->3,
  package_release_scripts 4->6 live sites; lane R moved the packing and
  signing checks from release.yml to package-release.yml (phase 4 home).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
- project_overview's watcher: mode always, interval and fix sentence only
  for polling/silent; probe counters and the path-bearing unprobed detail
  stay on code-index://stats and doctor (the daemon leg's saturated overview
  was 36 tokens over: an unbounded path on the first call). Test +
  mutation (act always true -> RED).
- bounding_site_registry: SERVED_PATHS_CAP registered (Intentional: bounds
  the check's cost, removes nothing from a reply).
- inline_cargo_build_registry: the release-only seam test skips a present
  Windows binary before its cargo build.
- wall_clock_registry: the two lane W timing tests are in ci-timing and out
  of ci.
- harness_settle_barrier: 62 leg-routed suites use the shared barrier.
- census_and_earned_zero_e2e: release.yml no longer spells package_digest
  at all; the fixture token is workflow_runs, and the prefix query is
  derived from it.
- owner_switch_rollback: the 'package owns nothing it claims' shape is now
  spelled [languages] rust = "builtin" (under auto rust is package-owned
  since the flip); the documented ownership mutation still turns all three RED.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Merge fix-0410 (gates the v0.41 full run found red) into train-0410
All checks were successful
Package CI / package (timeline) (pull_request) Successful in 31m5s
Package CI / package (xaml) (pull_request) Successful in 31m19s
Package CI / package (svelte) (pull_request) Successful in 32m24s
CI / CI lane wall-clock headroom (pull_request) Successful in 1m6s
Package CI / package (rust) (pull_request) Successful in 33m10s
Package CI / package (javascript) (pull_request) Successful in 33m19s
Package CI / package (ruby) (pull_request) Successful in 33m29s
Package CI / package (php) (pull_request) Successful in 33m27s
CI / guest crates (fmt, clippy, doc) (pull_request) Successful in 1m41s
Package CI / package (csharp) (pull_request) Successful in 34m17s
Package CI / package (typescript) (pull_request) Successful in 34m40s
CI / cargo check (MSRV 1.98) (pull_request) Successful in 19m50s
Package CI / package (python) (pull_request) Successful in 39m45s
CI / cargo doc (intra-doc links) (pull_request) Successful in 9m42s
CI (Windows) / fmt + clippy + build + test (windows) (pull_request) Successful in 43m10s
CI / cargo deny (pull_request) Successful in 12m19s
CI / cargo test (abi, 32-bit + wasm32) (pull_request) Successful in 13m14s
CI / cargo check (windows-gnu) (pull_request) Successful in 16m12s
Package CI / package-corpus (php) (pull_request) Successful in 49m40s
Package CI / package-corpus (javascript) (pull_request) Successful in 50m2s
Package CI / package-corpus (python) (pull_request) Successful in 49m58s
Package CI / package-corpus (csharp) (pull_request) Successful in 50m24s
Package CI / package-corpus (rust) (pull_request) Successful in 50m9s
Package CI / package-corpus (typescript) (pull_request) Successful in 48m29s
Package CI / package-corpus (ruby) (pull_request) Successful in 55m33s
Package CI / package overhead budgets (nightly, one package at a time) (pull_request) Has been skipped
CI / cargo test (pull_request) Successful in 27m50s
CI / OSS corpus (tier 1) (pull_request) Successful in 45m51s
CI / cargo test (daemon transport) (pull_request) Successful in 11m22s
CI / Plugin path cost + pool throughput (nightly) (pull_request) Has been skipped
439aa9ed42
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
h-dv/code-index!389
No description provided.