Release 0.37.0: split CI per language package (#354), #355, #356 #360

Merged
buildagent merged 39 commits from train-037 into master 2026-10-13 00:21:17 +02:00
Member

The v0.37 train. Every lane ran targeted tests and its own mutations, and I verified each lane myself: re-ran its suites and ran one mutation of my own, which went RED each time. One full local gate run on this tree, now green.

Epic #354: fix and release one language package without the rest

  • A, inline gate. package-release runs the package's own grading in a leg job on the tagged commit. The 50-minute poll for a package-ci leg (run 1089) is gone, along with package_ci_gate.sh.
  • B, one grading script. .forgejo/scripts/package_leg.sh <name> <release|corpus|full> is used by package-ci (package (<n>) = release tier, plus a parallel package-corpus (<n>)) and by package-release.
    • Each tier does one cargo build. COSI_PLUGIN_HOST_PREBUILT removes the nested worker build that took 11m51 in run 1090.
    • Ruby release tier from a clean target: 143 s locally. Run 1090's ruby leg took 74 min in CI.
  • E, one commit. Every package-release job checks out github.sha, and package_pipeline_commit.sh refuses a tag that isn't that commit, before anything is graded or published (run 1102's shape). Graded in release_gate/ci_path_gating.
  • D, catalog. A core release lists a package's own-release version only once the published channel serves it at the same digest; it names everything it withholds. Before publishing, it fetches every own-release row and every row the seed adds. The release notes give languages owner <lang> package for every language package.
  • C, ownership.
    • All seven language packages now compile their own walker copy under crates/guest/<pkg>/src/. JavaScript and TypeScript each have their own copy.
    • Every builtin walker is frozen (builtin_frozen), and every walker is gated to its own package's lanes.
    • Ownership, frozen and walker-parity suites are list-driven (support/owned.rs).
    • Measured: an edit to one walker moves only that package's bytes. Rebuilding all 9 extractors on this tree is byte-identical to the checked-in artifacts.
    • Walker parity matches the builtin exactly on every pinned corpus repo: sinatra 156, guzzle 131, flask 83, dapper 157, ripgrep 110, express 141, zod 401 files.

Field fixes (zammad-dev)

  • #355. Influence components carry package and component_lang beside the handle. search_text lang=typescript admits .vue/.svelte files by the language their regions contain (region_lang), and the imports half of search_symbols filters on the row's language.
  • #356. changed_symbols and review_diff take path_glob and page by file with a stale-checked cursor. On an 80-commit range, one 95k-char reply that never probed 243 of 284 files became complete, bounded pages.

Versions

  • Core 0.37.0.
  • php, python, csharp, rust, javascript and typescript are bumped to 0.1.3 (the ownership gating moved their bytes, by layout only) and ship with this release.
  • de.h-dv.ruby 0.10.1 is on its own release: withheld by lane D until its pkg/ release runs. That is the F drill.
  • tests/corpus/ruby-package-cost.json was re-measured for 0.10.1 with operator approval: +220 vm_step, isolated window.
  • baseline.json and cost-baseline.json are unchanged.

Full local run (this tree)

  • Green: fmt, clippy (workspace, first-party-tests, guest), rustdoc with --document-private-items -D warnings, the workspace nextest run (5091/5091), the daemon leg, timing (34/34) and guest gates.
  • Corpus: the CI corpus job (11 suites), guest-kit psr4_vs_builtin, tier 3, and precision_gate (7/7, phantom 0).
  • package_leg.sh <pkg> full for all 11 packages.
  • The first pass found two gates no lane ran (ci-timing release_gate_e2e, and the host_reads_no_live_package count). Both are fixed in 36453952, and ws/timing were re-run green.

Closes #355, #356. Part of #354 (F, the drill, runs after release).

🤖 Generated with Claude Code

https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj

The v0.37 train. Every lane ran targeted tests and its own mutations, and I verified each lane myself: re-ran its suites and ran one mutation of my own, which went RED each time. One full local gate run on this tree, now green. ## Epic #354: fix and release one language package without the rest - **A, inline gate.** package-release runs the package's own grading in a `leg` job on the tagged commit. The 50-minute poll for a package-ci leg (run 1089) is gone, along with `package_ci_gate.sh`. - **B, one grading script.** `.forgejo/scripts/package_leg.sh <name> <release|corpus|full>` is used by package-ci (`package (<n>)` = release tier, plus a parallel `package-corpus (<n>)`) and by package-release. - Each tier does one cargo build. `COSI_PLUGIN_HOST_PREBUILT` removes the nested worker build that took 11m51 in run 1090. - Ruby release tier from a clean target: 143 s locally. Run 1090's ruby leg took 74 min in CI. - **E, one commit.** Every package-release job checks out `github.sha`, and `package_pipeline_commit.sh` refuses a tag that isn't that commit, before anything is graded or published (run 1102's shape). Graded in `release_gate`/`ci_path_gating`. - **D, catalog.** A core release lists a package's own-release version only once the published channel serves it at the same digest; it names everything it withholds. Before publishing, it fetches every own-release row and every row the seed adds. The release notes give `languages owner <lang> package` for every language package. - **C, ownership.** - All seven language packages now compile their own walker copy under `crates/guest/<pkg>/src/`. JavaScript and TypeScript each have their own copy. - Every builtin walker is frozen (`builtin_frozen`), and every walker is gated to its own package's lanes. - Ownership, frozen and walker-parity suites are list-driven (`support/owned.rs`). - Measured: an edit to one walker moves only that package's bytes. Rebuilding all 9 extractors on this tree is byte-identical to the checked-in artifacts. - Walker parity matches the builtin exactly on every pinned corpus repo: sinatra 156, guzzle 131, flask 83, dapper 157, ripgrep 110, express 141, zod 401 files. ## Field fixes (zammad-dev) - **#355.** Influence components carry `package` and `component_lang` beside the handle. `search_text lang=typescript` admits `.vue`/`.svelte` files by the language their regions contain (`region_lang`), and the imports half of `search_symbols` filters on the row's language. - **#356.** `changed_symbols` and `review_diff` take `path_glob` and page by file with a stale-checked cursor. On an 80-commit range, one 95k-char reply that never probed 243 of 284 files became complete, bounded pages. ## Versions - Core 0.37.0. - php, python, csharp, rust, javascript and typescript are bumped to 0.1.3 (the ownership gating moved their bytes, by layout only) and ship with this release. - de.h-dv.ruby 0.10.1 is on its own release: withheld by lane D until its `pkg/` release runs. That is the F drill. - `tests/corpus/ruby-package-cost.json` was re-measured for 0.10.1 with operator approval: +220 vm_step, isolated window. - `baseline.json` and `cost-baseline.json` are unchanged. ## Full local run (this tree) - Green: fmt, clippy (workspace, first-party-tests, guest), rustdoc with `--document-private-items` `-D warnings`, the workspace nextest run (5091/5091), the daemon leg, timing (34/34) and guest gates. - Corpus: the CI corpus job (11 suites), guest-kit `psr4_vs_builtin`, tier 3, and `precision_gate` (7/7, phantom 0). - `package_leg.sh <pkg> full` for all 11 packages. - The first pass found two gates no lane ran (ci-timing `release_gate_e2e`, and the `host_reads_no_live_package` count). Both are fixed in 36453952, and `ws`/`timing` were re-run green. Closes #355, #356. Part of #354 (F, the drill, runs after release). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
A 46-commit field range returned ~112k chars even concise. Measured on an
80-commit range of this repository (284 files): the old single answer was
95,400 chars concise / 172,325 detailed, 70.7% symbol rows and 24.5% file
envelopes, and it showed 300 of 1,232 symbols while reporting
omitted_symbols: 9 (243 of 284 listed files were never probed).

- path_glob: SQLite's own GLOB operator on an in-memory connection, the
  dialect of every other path_glob; a rename matches on either side; the
  reply carries filter{matched,excluded} and a never-masked note.
- Paging by file: a page holds at most CHANGED_SYMBOLS_PAGE_BUDGET (100)
  symbols and CHANGED_FILES_PAGE_MAX (100) files and never splits a file;
  next_cursor/page_offset/remaining_files are always serialised.
- The cursor binds the call (tool, project, base, staged, path_glob) and
  the diff (sha256 over the resolved base commit, staged and every paged
  file's change + hunks); each mismatch is refused with its own hint.
- review_diff pages the same way; coverage_capped names the page; concise
  bounds test_role_files (12.9k of a 27.7k concise review) to 20 with
  test_role_files_total/omitted.
- context_pack keeps its pre-paging 300-symbol seed pool.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The field report: `de.h-dv.vue/typescript` and `de.h-dv.vue/vue` in
`resolution_gaps.dynamic_influence.components` and on find_callers /
find_references / find_callees rows (`influence.component`) read as
LANGUAGES. Neither is stored: both are assembled as
`package_key || '/' || component`, a component HANDLE. For a region
component the handle's local half is the region's language by
convention only, so `de.h-dv.vue/typescript` is spelled like a package
language id and means "the TypeScript regions the Vue package
extracted", whose rows are `typescript`.

1. The handle stays (it is the stable identity that tests, grants and
   the plugin bench pin), and the two facts it packs now ride beside it,
   READ FROM THE COMPONENT ROW, never split out of the string:
   * `RefInfluence` gains `package` and `component_lang`
     (`extraction_components.lang`), via one `ComponentLabel` map;
   * `DynamicInfluence` gains `component_labels`
     (`{component, package, component_lang}` per handle, same order as
     `components`), and its `semantics` says handles are not languages.
   Additive and serde-defaulted in both skew directions: an old daemon
   leaves the fields absent, an old client ignores them.
   In `concise`, `package` is dropped (the handle's part before its last
   `/`) and `component_lang` is kept only where it differs from the
   handle: `de.h-dv.xaml/xaml` already IS a language id, and repeating
   it cost agent_task_plugin_bench's plugin-wpf band 61 tokens into its
   reserve (12472 -> 12533). With the rule it passes unchanged.

2. A `lang` filter now ADMITS a file by the language it contains.
   `search_symbols` (s.lang) and `resolution_gaps` (r.lang) already
   filtered ROWS and found region code under `lang="typescript"`;
   `search_text` read `files.lang` alone and silently missed every
   `.vue`/`.svelte` script region, and `search_symbols`' imports arm read
   `f.lang` too, so a `<script lang="ts">` import was missed AND labelled
   `de.h-dv.vue/vue`. Chosen: MATCH, not disclose, because the
   row-level surfaces already define `lang="typescript"` as "TypeScript
   code", and a file-level surface disagreeing with them is the defect.
   The rule is ONE structural fact for every region host: a file matches
   language L when its own language is L OR a LIVE contribution to it
   was written by a component whose rows are L
   (`extraction_components.lang`; `ensure_delegated_component` holds one
   language per component). It stays honest: the hit keeps the file's
   own `lang` and names the admitting region language in a new
   `region_lang` (detailed and concise), whose doc says the census is
   the whole file's, so a matched line may lie outside the region.
   The imports arm now reads and reports `i.lang` (m0044), as the
   symbols arm reads `s.lang`.

NOT changed: `files.lang = 'de.h-dv.vue/vue'` on list_files /
file_outline / project_overview. That IS the Vue language's id: `vue`
is not a reserved local id, so `langid::row_lang` keeps it qualified,
exactly as `de.h-dv.xaml/xaml`.

Tests (daemon e2e through the real binary, frozen vue and svelte
packages), each mutation run RED:
* vue_package_e2e::an_influence_component_names_its_package_and_its_language_apart
* vue_package_e2e::a_lang_filter_admits_a_file_by_the_language_it_contains
* svelte_package_e2e::a_svelte_region_is_typescript_by_its_language_and_its_package_is_named_apart
* server concise_influence_tests + concise_text_hit_flattens_and_bounds_snippet
like_escape_tests' hand-built `imports` table gains the `lang` column
every migrated index has had since m0044.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A row on pkg%2F<id>%2Fv<version> is owned by package-release. v0.36.0 was
cut between the ruby 0.10.0 bump commit and that version's own release,
and both its signed catalog and its channel seed listed a URL nothing
served.

- release_registry.sh: synthesize/stage/audit take PACKAGE_CHANNEL (the
  published pkg-channel registry.v2.json, or `-`; unset is a refusal).
  An own-release version is kept only when that channel lists it at the
  same package_digest; every other one is withheld, named and warned.
  The audit re-derives with the same rule and channel. The channel
  decides membership only: every kept value is still the source's.
- build_channel: a release catalog (an include, or the first input with
  no base) never ADDS an own-release version the base does not list and
  this run does not release; reported as merge_withheld.
- package_channel_check.sh: --downloadable-new <catalog> <base|-> and
  --downloadable-own <catalog> fetch every added / own-release row.
- release.yml: the release job fetches the published channel before
  staging and fetches the catalog's own-release rows before publishing;
  the seed checks every row it adds before publishing (and again on the
  re-read), plus channel-vs-core consistency before publishing.
- release notes: every language package block gives
  `code-index languages owner <lang> package`; under the default `auto`
  owner `languages enable` alone leaves the builtin owning it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ONE TABLE, ONE IMPLEMENTATION (support/owned.rs, support/walker_parity.rs):
- OWNED lists each language whose package owns its extractor source
  (package, lang, builtin id, owned walker(s), frozen builtin, deltas
  file, pinned corpus repo). Ruby is row 1, PHP row 2.
- The #345 Ruby-only suites are now generic functions over a row:
  the table is the tree (deltas files == rows == owned modules lib.rs
  compiles; each deltas file names its row's builtin), the walker is an
  input of its package only, a walker edit moves only that package's
  record, no host code compiles the walker. <pkg>_source_ownership.rs
  names its row; the walker parity body is include!d by
  <pkg>_walker_parity.rs, which supplies only the native walker fn and
  reads COSI_CORPUS_DIR itself (so package-ci's corpus step and
  corpus_require_floor still see it). Fixture population = the package's
  plugin.toml [[fixtures]]; module facts are compared where the builtin
  states them.
- crates/plugins/tests/builtin_frozen.rs (was ruby_builtin_frozen.rs):
  FROZEN pins every frozen builtin once, with its owning packages (one
  builtin may be copied by two packages: javascript/typescript), and is
  held equal to what crates/guest/*/parity-deltas.toml measure against.

EVERY WALKER IS COMPILED INTO ITS OWN PACKAGE'S LANES ONLY:
- build.rs sets code_index_guest_<package> per lane from a lane->package
  table (jsx -> javascript, tsx -> typescript); lib.rs gates csharp, php,
  python, rust and typescript (javascript+typescript) like ruby; wasm.rs
  gates each dispatch arm.
- MEASURED (build.py, all nine artifacts): gating ONLY php out of the
  other lanes changed EVERY other package's bytes (same strings, laid out
  in another order; code section same size, data -8..+8 bytes), so any
  ungated walker ties every package to it. With every walker gated, a
  later ownership move changes only its own package: simulated here for
  python (copy) and javascript (split from typescript) -> only python,
  javascript and jsx moved; csharp, php, ruby, rust, typescript, tsx
  byte-identical.
- inputs.rs: guest_modules() reads each #[path] module's gate (one line
  or rustfmt-wrapped; any other cfg shape refused); a builtin walker is a
  shared input of the packages its cfg names, an owned walker must be
  gated to exactly its owner.

PHP OWNS ITS WALKER: crates/guest/php/src/php.rs (the builtin minus its
unit tests, verbatim but for its header), build.sh, parity-deltas.toml
with no delta (MEASURED: fixtures 8 files and php-guzzle 131 files equal);
php_walker_parity also takes over the PHP cases shared_guest_parity
graded (tests/fixtures/php + the malformed case), which now skips the
OWNED ids (43 files, five languages). ci_path_gating: php rows; the
shared-input premise moves to common.rs, and a builtin walker routes to
its own leg only.

The package records are stale until the next commit rebuilds them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- diff_paging_e2e.rs (both legs): pages cover the diff exactly once
  against git's file list and the fixture's 400 functions; path_glob
  matches what list_files matches for six globs; a filter that empties
  the diff never reads 'No changes'; brace groups refused on both tools;
  a cursor is refused as foreign (other glob, other tool) or as stale
  (an edit moved a hunk), with different hints; review_diff pages the
  same diff and names the page in coverage_capped; concise review lists
  20 of 25 test-role files with total/omitted.
- unit: diff stamp moves with every input it covers; cursor round-trip
  and refusal by cause; the matcher is SQLite GLOB; page fields always
  serialised.
- bounding_site_registry rows for CHANGED_SYMBOLS_PAGE_BUDGET,
  CHANGED_FILES_PAGE_MAX, CONCISE_TEST_ROLE_FILES_MAX.
- changed_symbols_bounds_large_diff asserts the paging contract.
- The new arguments pushed the startup payload 112 tokens into its
  reserve; changed_symbols' description restated its base/staged
  parameter docs and two parentheticals — trimmed, net tools/list
  61,087 -> 61,044 chars.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
B: .forgejo/scripts/package_leg.sh <name> <release|corpus|full> is the ONE
implementation of a package's grading. package-ci.yml's `package (<name>)`
runs the release tier and a new `package-corpus (<name>)` runs the corpus
tier beside it (same matrix, same fail-open scope), so a PR waits for the
longer half, not the sum. Each tier compiles its graph in ONE cargo build
that names every test target it runs and the worker binary; the fixtures
take that worker under COSI_PLUGIN_HOST_PREBUILT=1 instead of rebuilding it
as `-p code-index-plugin-host` alone (11m51 of run 1090's ruby leg).
Measured locally, ruby release tier from an empty target dir: 291 s before,
143 s after (under a heavier machine load).

A: package-release.yml's `ci-gate` (poll /actions/tasks for another run's
leg under a 50-minute deadline; run 1089) is replaced by `leg`, which runs
package_leg.sh at the release tier on the tagged commit itself. It holds
no secret; every later job needs it; no skip input remains.
package_ci_gate.sh and require_ci_green.sh's NOT_RED_JOBS/REQUIRED_EVENTS
arms, used only by it, are gone.

E: every package-release job checks out github.sha and runs
package_pipeline_commit.sh, which refuses unless the tag names that same
commit (annotated tags and a tag-object GITHUB_SHA peeled). Run 1102's
shape (master's workflow, the tag's older scripts) now stops in `leg`,
before anything is packed or published.

Tests grade the new shape: release_gate::a_package_release_grades_its_package_inline,
release_gate::a_package_release_runs_one_commits_pipeline,
ci_path_gating::the_corpus_legs_mirror_the_package_legs and
ci_path_gating::script::a_package_release_runs_only_its_tags_own_commit;
the leg readers (ci_path_gating, corpus_require_floor,
ignored_test_reachability, release_gate_e2e) read the script.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The HEAD == GITHUB_SHA arm of package_pipeline_commit.sh had no case of
its own: the run-1102 case trips both arms. A tag-vs-HEAD mutation
SURVIVED (behind that arm the two comparisons are one), so the case now
checks out master under a workflow that IS the tag's commit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The previous commit compiles every walker into its own package's lanes
only, and the PHP lane from the package's own walker. That moves the bytes
of all nine build.py artifacts (MEASURED, crates/guest/languages/build.py,
every lane; the tree before reproduced all nine checked-in artifacts byte
for byte):

  php        1875293 -> 1873433   python     1958546 -> 1957667
  rust       1894214 -> 1893866   csharp     1896625 -> 1894787
  javascript 1891349 -> 1890119   jsx        1894638 -> 1893400
  typescript 1894165 -> 1892919   tsx        1894638 -> 1893400
  ruby       1899124 -> 1896865

Why every package and not only php: with all walkers compiled into a
lane, dropping a DEAD one still reorders the surviving constants in the
data segment (same strings, other order) and with them the code. Gating
php alone moved all eight others; gating all of them at once costs this
one rebuild, and afterwards an ownership move of one language moves only
that language's package (measured on a python and a javascript
simulation, see the previous commit).

- extractor.wasm (+ jsx/tsx) rebuilt by build.py; artifact rows follow
  the bytes; patch version +1 (php, rust, python, csharp, javascript,
  typescript 0.1.2 -> 0.1.3; ruby 0.10.0 -> 0.10.1).
- .digest, the generated input rows of .provenance and the catalog by
  `package_records -- --write` (it re-recorded exactly these seven; svelte,
  timeline, vue, xaml unchanged). Every builtin walker left the other
  packages' input sets; php records its own walker and build.sh. Catalog
  rows stay on the release class they name (ruby on its own
  pkg%2F…%2Fv0.10.1, the six others on the host segment).
- Facts: the walkers' sources are unchanged (php's is the builtin's,
  measured equal by php_walker_parity on its fixtures and php-guzzle);
  every leg's own-bytes step (COSI_GUEST_REBUILD=1, fixture expectations
  through the new extractors) and its four release gates are green.

ruby_package_cost pins package_version 0.10.0 in
tests/corpus/ruby-package-cost.json and refuses the bump until the
operator re-records it; not blessed here.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The PostToolUse hook asked changed_symbols for the whole uncommitted
  diff and searched it for the edited file; a paged diff answers with
  page one, so the edited file was found only on small diffs. It now
  asks for the edited file by path_glob (GLOB metacharacters escaped as
  classes). New test: 120 modified files sort before the edited one,
  putting it on page two; the hook still reports widen.
- harness_settle_barrier: diff_paging_e2e joins the leg-routed
  population served by the shared initialize barrier (55 -> 56).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/plugins/src/common.rs sits in the plugins crate's target tree AND
is compiled into crates/guest/languages by #[path]. The test attributed it
to the plugins crate only, so the PHP package's own walker
(crates/guest/php/src/php.rs, compiled solely by the languages guest)
resolving 5 edges into common.rs read as an undeclared
languages -> plugins crossing. A file now belongs to the crate whose
target tree holds it AND to every crate whose #[path] compiles it; an edge
is cross-crate only when no crate compiles both ends.

RED before this change (the PHP ownership tree, unchanged test):
"crates/guest/languages -> crates/plugins (5 edge(s)):
crates/guest/php/src/php.rs -> crates/plugins/src/common.rs". Green after.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The count was derived from changed.files, which is one PAGE now: a
git rm --cached file on another page would have been counted as an
untracked file the staged review skipped. collect_changed_once counts it
once over the whole (glob-filtered) diff before paging and the cap.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
For the next lanes:
- support::owned::STILL_ON_THE_FROZEN_BUILTIN names (builtin, package)
  for a package still compiling a builtin another package already owns a
  copy of: javascript and typescript share crates/plugins/src/typescript.rs
  and move in two changes. Empty today; every other package is held off
  every frozen builtin, and each entry is held TRUE.
- builtin_frozen: a builtin with a module directory of its stem
  (python.rs + python/) is frozen whole: every file there needs its own
  FROZEN row with the same owners.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
de.h-dv.csharp now builds from crates/guest/csharp/src/csharp.rs, a copy
of the builtin crates/plugins/src/csharp.rs minus its two
#[cfg(all(test, …))] modules (tests, qualifier_tests), verbatim but for
its header (diff: only the 20 header lines and the 1458 test lines
differ). The builtin is frozen (builtin_frozen FROZEN row).

- crates/guest/languages/src/lib.rs: the csharp #[path] points at the
  package's walker; its code_index_guest_csharp gate is kept.
- crates/guest/csharp/: build.sh (build.py csharp), parity-deltas.toml
  with no delta against the builtin's sha256.
- support/owned.rs: the csharp row (corpus repo cs-dapper).
- csharp_source_ownership.rs, csharp_walker_parity.rs: the shared suites
  over the csharp row; walker parity also takes over the host C# fixtures
  (tests/fixtures/csharp, 6 files) plus a malformed case. MEASURED equal:
  fixtures 6 files, cs-dapper 157 files, no delta.
- shared_guest_parity: csharp joins OWNED and leaves the walk; the floor
  is the MEASURED 37 files over four languages (43 before).
- ci_path_gating: csharp rows for the walker, deltas, build.sh and
  csharp_walker_parity.

The package records are stale until the next commit rebuilds them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
build.py over ALL nine lanes into a temp --output, cmp against
tests/packages: ONLY csharp/extractor.wasm moved (1894787 -> 1894779,
the walker's path is the one source string that changed); javascript,
jsx, php, python, ruby, rust, typescript, tsx byte-identical.

- csharp/extractor.wasm and its provenance artifact row follow the bytes.
- package_records -- --write re-recorded csharp only (digest,
  provenance inputs: the package now records its own walker and
  build.sh, and no longer the builtin; catalog row 0.1.3 digests).
  The other six build.py packages report lib.rs as a SHARED move and are
  left alone, by design.
- No version bump: 0.1.3 is already this train's unreleased bump.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/guest/javascript/src/typescript.rs is the builtin
crates/plugins/src/typescript.rs minus its two #[cfg(all(test, ...))]
modules, verbatim but for a header prepended above the builtin's module
doc (diff: 0a1,22 and the test modules only). lib.rs compiles it as
`javascript` into the javascript lanes only (javascript, jsx); the
builtin's gate narrows to typescript; wasm.rs's javascript|jsx arm calls
crate::javascript. build.rs unchanged.

- build.sh (build.py javascript jsx), parity-deltas.toml against the
  builtin's sha256 with no delta (MEASURED equal: package fixtures 7
  files, host fixtures 7 + the .jsx case, js-express 141 files / 1917
  symbols / 20365 refs / 388 imports).
- FROZEN pins typescript.rs (owners: javascript); OWNED gains the
  javascript row (builtin_id "typescript", lang "javascript",
  js-express); STILL_ON_THE_FROZEN_BUILTIN names (typescript.rs,
  typescript) until the typescript move.
- tables/tsx.rs STAYS SHARED and is now a declared, tested fact:
  owned::SHARED_TABLES + assert_every_table_has_its_declared_packages
  (it is the pinned tsx grammar's generated kind table, not extractor
  source; both the jsx and the tsx lane parse with that one grammar).
- javascript_source_ownership, javascript_walker_parity (also takes over
  the host's JavaScript fixtures and .jsx case). shared_guest_parity now
  skips owned LANGUAGES (lang_for_path), not plugin ids, since one
  builtin plugin extracts both; floor 43 -> 36 (MEASURED 36).
- ci_path_gating: javascript rows; crates/plugins/src/typescript.rs is
  now recorded by the typescript leg only.

Rebuild (build.py, all nine lanes into a temp --output, cmp against
tests/packages): ONLY javascript/extractor.wasm and jsx-extractor.wasm
moved (same size, 14 bytes: the #[path] string of the walker's panic
locations); csharp, php, python, ruby, rust, typescript, tsx identical.
No version bump (C0 already bumped 0.1.3 in this train). Artifact rows
and notes updated; package_records --write re-recorded javascript only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/guest/python/src/python.rs + python/{scope,module_facts}.rs: the
builtin crates/plugins/src/python.rs and its module directory, verbatim but
for a header on python.rs and without their #[cfg(all(test, ..))] modules
(diff-checked: module_facts.rs byte-identical). lib.rs compiles it into the
python lane only, under the gate C0 put on every walker.

- build.sh, parity-deltas.toml (builtin python.rs + its sha256), no delta:
  MEASURED equal on the 9 fixtures and on all 83 python-flask files.
- builtin_frozen: FROZEN rows for python.rs, python/module_facts.rs and
  python/scope.rs (owner python).
- owned.rs: the python row. Each row now carries corpus_files_floor (the
  measured population of its pinned repo: sinatra 156, guzzle 131, flask
  83): the shared 100-file floor refused python-flask with every file equal.
- python_source_ownership / python_walker_parity; the latter takes over the
  Python cases shared_guest_parity graded (tests/fixtures/python, the
  __all__ / metadata sources, the malformed def). shared_guest_parity skips
  python (34 files, four languages, floor re-measured); its builtin-only
  module_bindings assertions stay on the host.
- ci_path_gating: python package rows; the per-walker shared-input example
  moves to csharp.rs; frozen php.rs/python.rs/python/scope.rs route to no leg.
- manifest_layering: #[path] owners propagate transitively, so
  python/scope.rs compiled by the #[path]-compiled guest python.rs belongs
  to the languages guest (RED before: both files unattributed).

The python package record is stale until the next commit rebuilds it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
build.py over ALL nine lanes into a temp output, cmp against
tests/packages: only python/extractor.wasm moved (1957667 -> 1957651
bytes; the walker's source path is spelled in its panic locations);
csharp, javascript, jsx, php, ruby, rust, typescript, tsx byte-identical.
No version bump: C0 already bumped python in this unreleased train.

The artifact row follows the bytes; package_records -- --write
re-recorded python only (its own walker, module directory and build.sh
replace the builtin's shared inputs) and the catalog's python digests;
check mode is clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
RUST OWNS ITS WALKER, by lane C0's recipe:
- crates/guest/rust/src/rust.rs: the builtin crates/plugins/src/rust.rs
  (lines 1-2228) minus its two #[cfg(all(test, ...))] modules, verbatim
  under a 19-line header (diff: only the header and the test modules).
- crates/guest/languages/src/lib.rs repoints the rust #[path] to it; the
  code_index_guest_rust gate stays. build.rs unchanged.
- crates/guest/rust/build.sh (PHP's, for the rust lane) and
  parity-deltas.toml (builtin crates/plugins/src/rust.rs at its sha256,
  no delta).
- FROZEN row for rust.rs in crates/plugins/tests/builtin_frozen.rs; OWNED
  row in support/owned.rs (corpus repo rust-ripgrep).
- rust_source_ownership.rs and rust_walker_parity.rs. The parity target
  also takes over what shared_guest_parity graded for Rust:
  tests/fixtures/rust (12 files) + the malformed case, the walker half of
  the missing-node cases, and the two #348 binding-qualifier wire tests.
  MEASURED equal: package fixtures 12 files, host fixtures 12 + 1, and
  rust-ripgrep 110 files / 4872 symbols / 43845 refs / 1046 imports.
- shared_guest_parity: rust joins OWNED; the fixture floor is re-measured
  at 31 (43 - Rust's 12, four languages). The host serializer's
  missing-node flag check stays on the host (the walker compare moved).
- ci_path_gating: rust rows (walker, deltas, build.sh, both new targets);
  the frozen rust.rs is now recorded by no package, so the shared-input
  premise grades python.rs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
build.py over ALL lanes into a temp --output, cmp against tests/packages:
ONLY rust/extractor.wasm moved (1893866 -> 1893858 bytes); the eight
other build.py artifacts and every grammar are byte-identical. The 8
bytes are the embedded source path: rustc's panic location for the
walker reads crates/guest/languages/src/../../rust/src/rust.rs instead of
.../../../plugins/src/rust.rs; the walker's code is the builtin's.

- extractor.wasm and its artifact row; the provenance note names the
  package-owned walker. No version bump: C0 already moved rust to 0.1.3
  in this unreleased train.
- package_records -- --write re-recorded rust only (its input rows: its
  own walker and build.sh in, the frozen builtin out); the catalog row
  follows its digest. The other packages report the lib.rs shared input
  as moved and keep their records (their bytes did not move); check
  mode exits 0.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
crates/guest/typescript/src/typescript.rs is the builtin
crates/plugins/src/typescript.rs minus its two #[cfg(all(test, ...))]
modules, verbatim but for its header (diff: 0a1,23 and the test modules
only; its body is identical to javascript's copy). lib.rs's `typescript`
module now points at it, gated to the typescript lanes (typescript, tsx);
wasm.rs and build.rs unchanged.

- build.sh (build.py typescript tsx), parity-deltas.toml against the
  builtin's sha256 with no delta (MEASURED equal: package fixtures 9
  files, host fixtures 9 + the .tsx case + the three typeof cases, ts-zod
  401 files / 9667 symbols / 92713 refs / 1556 imports).
- FROZEN typescript.rs owners: javascript, typescript. OWNED gains the
  typescript row (ts-zod); STILL_ON_THE_FROZEN_BUILTIN is empty again.
- typescript_source_ownership, typescript_walker_parity (takes over the
  host's TypeScript fixtures, the .tsx case and the guest half of the
  typeof nested-member cases; their native assertions stay in the host).
  shared_guest_parity: typescript is an owned language, its unreachable
  dispatch arm is gone; floor 36 -> 27 (MEASURED 27).
- ci_path_gating: typescript rows; crates/plugins/src/typescript.rs is
  recorded by no package any more.
- tables/tsx.rs stays shared by javascript and typescript, declared in
  owned::SHARED_TABLES and asserted by both ownership suites.

Rebuild (build.py, all nine lanes into a temp --output, cmp against
tests/packages): ONLY typescript/extractor.wasm and tsx-extractor.wasm
moved (same size, 14 bytes each: the walker's #[path] string);
javascript, jsx, csharp, php, python, ruby, rust identical. No version
bump (0.1.3 is this train's). Artifact rows and notes updated;
package_records --write re-recorded typescript only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
#	crates/indexer/tests/ci_path_gating.rs
#	crates/plugins/tests/builtin_frozen.rs
#	crates/plugins/tests/shared_guest_parity.rs
# Conflicts:
#	crates/indexer/tests/ci_path_gating.rs
#	crates/plugins/tests/builtin_frozen.rs
#	crates/plugins/tests/shared_guest_parity.rs
# Conflicts:
#	crates/indexer/tests/ci_path_gating.rs
#	crates/plugins/tests/builtin_frozen.rs
#	crates/plugins/tests/shared_guest_parity.rs
The python lane made the corpus floor a per-row field; the four lanes merged
after it predate the field. Each floor is the count that lane measured on its
pinned repo, re-measured on the integrated tree: cs-dapper 157, rust-ripgrep
110, js-express 141, ts-zod 401.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
0.10.1 is a layout-only rebuild from #354's per-package walker gating.
Isolated window (load 0.2): package vm_step 27929862 -> 27930082 (+220),
fullscan/sort/autoindex unchanged, wall ratio 131%. Walker parity on
ruby-sinatra is identical. baseline.json and cost-baseline.json untouched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
train-037: two gates the lanes' targeted runs did not reach
Some checks failed
CI / cargo check (windows-gnu) (pull_request) Failing after 18m15s
Package CI / package (timeline) (pull_request) Successful in 29m38s
Package CI / package (xaml) (pull_request) Successful in 29m53s
Package CI / package (php) (pull_request) Successful in 30m19s
CI / CI lane wall-clock headroom (pull_request) Successful in 57s
Package CI / package (svelte) (pull_request) Successful in 30m46s
Package CI / package (vue) (pull_request) Successful in 30m59s
Package CI / package (rust) (pull_request) Successful in 31m27s
CI / guest crates (fmt, clippy, doc) (pull_request) Successful in 1m34s
Package CI / package (csharp) (pull_request) Successful in 31m31s
Package CI / package (javascript) (pull_request) Successful in 31m41s
Package CI / package (typescript) (pull_request) Successful in 31m52s
Package CI / package (ruby) (pull_request) Successful in 32m1s
CI / cargo test (abi, 32-bit + wasm32) (pull_request) Successful in 16m15s
CI / cargo deny (pull_request) Successful in 12m41s
Package CI / package (python) (pull_request) Successful in 37m55s
CI / cargo doc (intra-doc links) (pull_request) Successful in 9m2s
CI / cargo check (MSRV 1.98) (pull_request) Successful in 12m35s
CI / cargo clippy (pull_request) Successful in 12m58s
Package CI / package-corpus (python) (pull_request) Successful in 47m2s
Package CI / package-corpus (javascript) (pull_request) Successful in 47m19s
Package CI / package-corpus (csharp) (pull_request) Successful in 47m20s
Package CI / package-corpus (php) (pull_request) Successful in 47m24s
Package CI / package-corpus (rust) (pull_request) Successful in 47m22s
Package CI / package-corpus (typescript) (pull_request) Successful in 45m17s
Package CI / package-corpus (ruby) (pull_request) Successful in 52m26s
CI / cargo test (daemon transport) (pull_request) Has been cancelled
CI / Plugin path cost + pool throughput (nightly) (pull_request) Has been cancelled
CI / OSS corpus (tier 1) (pull_request) Has been cancelled
CI / cargo test (pull_request) Has been cancelled
36453952a7
* release_gate_e2e (ci-timing profile) asserts the influence block as an
  exact object; #355 added package and component_lang beside the handle,
  so the expectation now names both (fx::XAML_PKG_ID, fx::XAML_LANG).
* host_reads_no_live_package: release_gate.rs gained one live read in
  lane D, the release-notes test that walks tests/packages to name every
  language package's owner switch — the same release-pipeline purpose as
  the row's other 29, so the exact count is 30 with the reason named.

Both were red in the one full run (ws, timing) and nowhere else.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ci_path_gating: leg_step is cfg(unix) like its only caller (PR #360 windows-gnu)
Some checks failed
Package CI / package (xaml) (pull_request) Successful in 29m53s
Package CI / package (svelte) (pull_request) Successful in 30m36s
Package CI / package (timeline) (pull_request) Successful in 30m40s
Package CI / package (php) (pull_request) Successful in 30m42s
Package CI / package (vue) (pull_request) Successful in 30m43s
Package CI / package (typescript) (pull_request) Successful in 30m51s
Package CI / package (rust) (pull_request) Successful in 31m7s
Package CI / package (csharp) (pull_request) Successful in 31m48s
Package CI / package (javascript) (pull_request) Successful in 31m52s
CI / CI lane wall-clock headroom (pull_request) Successful in 1m10s
CI / guest crates (fmt, clippy, doc) (pull_request) Successful in 1m34s
Package CI / package (ruby) (pull_request) Successful in 32m26s
Package CI / package (python) (pull_request) Successful in 38m0s
CI / cargo check (MSRV 1.98) (pull_request) Successful in 18m7s
CI (Windows) / fmt + clippy + build + test (windows) (pull_request) Successful in 39m27s
CI / cargo doc (intra-doc links) (pull_request) Successful in 9m26s
CI / cargo deny (pull_request) Successful in 11m5s
CI / cargo test (abi, 32-bit + wasm32) (pull_request) Successful in 12m7s
CI / cargo check (windows-gnu) (pull_request) Successful in 14m54s
Package CI / package-corpus (csharp) (pull_request) Successful in 47m49s
Package CI / package-corpus (php) (pull_request) Successful in 47m48s
Package CI / package-corpus (python) (pull_request) Successful in 47m51s
Package CI / package-corpus (javascript) (pull_request) Successful in 47m57s
Package CI / package-corpus (rust) (pull_request) Successful in 48m6s
Package CI / package-corpus (typescript) (pull_request) Successful in 46m18s
Package CI / package-corpus (ruby) (pull_request) Successful in 52m58s
CI / cargo test (pull_request) Failing after 28m7s
CI / cargo test (daemon transport) (pull_request) Has been skipped
CI / OSS corpus (tier 1) (pull_request) Successful in 42m21s
CI / Plugin path cost + pool throughput (nightly) (pull_request) Has been skipped
ea86a315cb
The pipeline lane's rewrite left script::the_leg_scope_check_fails_open,
inside the cfg(unix) module, as leg_step's only caller, so the helper was
dead code on Windows and -D warnings made both the windows-gnu clippy
check (run 1106) and the native Windows job (run 1105) red. Verified with
CI's own command locally: cargo clippy --workspace --all-targets --target
x86_64-pc-windows-gnu -- -D warnings is clean.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PR #360's CI red — reconnect_e2e "try 1: the successor serves", left
47137 (the leaving daemon) right 47666 (the successor) — failed after
5.49 s. Every try holds the leaving daemon's lock for 3 s, so two tries
cannot legitimately finish in under 6 s: the successor was never ready.
`Daemon::wait_ready` read the root's registration, which still named
the LEAVING daemon (lock held, registration published through its
drain), connected to its port and got a reconcile answer — from
whatever process the kernel had handed that closed loopback port to.
On a runner full of NO_AUTH test daemons, another test's daemon.

Not the startup beacon: the idle-exit seam is armed after
`StartupBeacon::finish` has joined the beat thread, in the same
sequential `run`, so no beat or final write can follow an idle exit.

Fix: a registration counts for readiness only when its pid is this
child's. A live process keeps its listener bound, so its own
registration's port cannot be recycled. The timeout message now also
names the last registration that belonged to another pid.

New: reconnect_e2e::wait_ready_does_not_accept_a_recycled_port_from_the_leaving_daemon
recycles the port on purpose (binds the leaving daemon's old port inside
the #350 window and forwards it to an unrelated daemon). RED on the old
harness with the CI signature; GREEN after.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Merge lane37-takeover-race into train-037
All checks were successful
Package CI / package (svelte) (pull_request) Successful in 29m57s
Package CI / package (xaml) (pull_request) Successful in 30m0s
CI / cargo clippy (pull_request) Successful in 24m5s
Package CI / package (timeline) (pull_request) Successful in 30m22s
Package CI / package (rust) (pull_request) Successful in 31m7s
CI / CI lane wall-clock headroom (pull_request) Successful in 1m7s
Package CI / package (php) (pull_request) Successful in 31m25s
Package CI / package (javascript) (pull_request) Successful in 32m0s
CI / guest crates (fmt, clippy, doc) (pull_request) Successful in 1m39s
Package CI / package (csharp) (pull_request) Successful in 32m9s
Package CI / package (typescript) (pull_request) Successful in 32m9s
Package CI / package (ruby) (pull_request) Successful in 32m59s
CI / cargo check (MSRV 1.98) (pull_request) Successful in 18m41s
Package CI / package (python) (pull_request) Successful in 38m48s
CI / cargo doc (intra-doc links) (pull_request) Successful in 9m29s
CI / cargo deny (pull_request) Successful in 11m34s
CI (Windows) / fmt + clippy + build + test (windows) (pull_request) Successful in 42m13s
CI / cargo test (abi, 32-bit + wasm32) (pull_request) Successful in 12m26s
CI / cargo check (windows-gnu) (pull_request) Successful in 15m50s
Package CI / package-corpus (csharp) (pull_request) Successful in 47m49s
Package CI / package-corpus (javascript) (pull_request) Successful in 47m53s
Package CI / package-corpus (python) (pull_request) Successful in 48m1s
Package CI / package-corpus (php) (pull_request) Successful in 48m12s
Package CI / package-corpus (rust) (pull_request) Successful in 48m12s
Package CI / package-corpus (typescript) (pull_request) Successful in 46m9s
Package CI / package-corpus (ruby) (pull_request) Successful in 53m18s
CI / cargo test (pull_request) Successful in 28m35s
CI / OSS corpus (tier 1) (pull_request) Successful in 42m53s
CI / cargo test (daemon transport) (pull_request) Successful in 12m59s
CI / Plugin path cost + pool throughput (nightly) (pull_request) Has been skipped
1750615288
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!360
No description provided.