Release 0.37.2: Rust module-alias calls (#367), scoped evidence_incomplete (#368), honest replies (#366), nightly CI green #373

Merged
buildagent merged 21 commits from train-0372 into master 2026-10-14 11:29:00 +02:00
Member

Fixes what dogfooding v0.37.1 on this repository found, and the red nightly CI. The version is bumped in this PR with version_bump.sh. Schema 81 → 82 (m0082: one full resolve on upgrade, no re-parse). Packages are unchanged.

#367: Rust calls through a workspace module alias resolve (lane E)

  • Cause: the alias expansion feeding temp.qual_groups (crates/indexer/src/index.rs) only admitted modules spelled ./crate::/self::/super::. use code_index_core::lang_profile as lp; lp::f() was never expanded, although the fully spelled path resolved.
  • Fix: an aliased module is also admitted when its head passes rust_use_head_is_workspace, the check follow_own_imports already uses. The same-file join, the "alias declared twice binds nothing" gate and whole-qualifier matching are unchanged. The lookup reads Rust imports only, so it costs nothing in other languages (fullscan_step is identical to master).
  • m0082 re-resolves once on upgrade. A schema-81 index of this repo migrates with parsed=0 and equals a fresh index bind for bind.
  • Measured, bind diff against master:
    • rust-analyzer +50 tier1q_alias, all hand-checked correct.
    • this repo +68, 40/40 sample correct.
    • the other 8 pinned repos identical; nothing lost or retargeted.
    • precision_gate 7/7, phantom 0, with a new Rust alias probe and decoy.
  • Re-records, all operator-approved:
    • tier3-baseline.json: rust-analyzer only, exactly the +50.
    • host-package-lifecycle.json: schema condition only, 0 counts moved.
    • ruby-package-cost.json: schema 82. vm_step +0.006%, inside the record's 0.02% determinism. Measured on one box: master 27,933,085 / 27,932,224 against this binary 27,935,458 / 27,934,443.
    • baseline.json and cost-baseline.json unchanged.

#368: evidence_incomplete only from partial files that could reference the target (lane P)

  • The single downgrade site (downgrade_absence_verdict) now counts a partial file only if its rows share the target's language (filter_admits), an active bridge leads into it, or its languages are unknown.
  • A package file that stopped before handing off its script regions counts for the host's route languages (TypeScript/JavaScript, from EMBEDDED_ROUTES). There is no path heuristic: a partial Rust file under tests/ still downgrades a Rust target.
  • Replies: partial_sources_in_scope (a count) sits beside the unchanged partial_sources_in_index. The resolution_gaps census states its scope and gives in_requested_scope when the call is filtered. The agent-task token cost is exactly master's (12,495).

Replies that state what they did not measure (lane D)

  • #366: defining_file_verified is set where the defining file is actually verified, for both check_rename and safe_delete. A full window gets a by-path pass in both tools.
  • changed_symbols / review_diff: a page with files remaining ships omitted_symbols_unmeasured: "paged_files_not_counted" instead of an unmeasured 0.
  • spellings_present: when it is capped, it says "the N nearest of the M".
  • search_text: a bare AND/OR/NOT/NEAR is an operator only when single terms strictly alternate. CREATE TABLE IF NOT EXISTS t was silently answered as a boolean (339 files); it is now one literal phrase (1). (a OR b) c still works. Behaviour change: an unparenthesised alpha OR beta gamma is now a literal phrase.
  • Recovery log: "cannot be recovered" names what moved (packages, bytes, grants), or says the record cannot tell (host or config).

Nightly CI (red since about 2026-10-06; fixed here)

  • bench_generation_gc (lane G): the bench promoted a hand-written, empty generation. 0.37.1's correct Switch::owed mark then re-extracted the files and emptied it, and the vacuity guard fired. The bench now builds the mirror generation for real. The guard is just as strict, and it still times 502/12,502 and 100,002/2,500,002 rows.
  • language_package_overhead (lane N): it compared native output against a binary pinned at 312a242e (2026-10-03). Every difference maps to an intended, blessed change: #309, #326, #335, #336, #337, #348, and the migration span fixes. Nothing was unexplained.
    • The check now grades the shipped no-package CLI against the in-process builtins.
    • Release-to-release drift stays owned by corpus_ratchet/corpus_stage; dropping a CLI builtin turns this check red.
    • The CI job no longer builds a second tree (−5.5 min).

Verification (I re-ran each lane's tests and at least one mutation myself; all RED, restored by md5)

  • One full local run, plus every nightly-only CI step locally on the final tree 9a89ac6a:
    • fmt; clippy (workspace, first-party, windows-gnu, guest); rustdoc --document-private-items -D warnings;
    • workspace nextest 5137/5137, daemon leg 942/942, ci-timing 34/34;
    • the CI corpus job (11 suites), psr4_vs_builtin, tier 3, corpus_scale, mixed-load, overview_scale_2m;
    • precision_gate 7/7, phantom 0;
    • all 11 package_leg.sh <pkg> full and all 7 overhead legs. The js leg was re-measured quiet after a contended run at load 10.
    • The full plugin-path-cost job: phase2 acceptance (7 languages, signed install, MCP, watch, restart), language_package_overhead, plugin_path_cost, package_pool, grammar_ab, claim_domain_scale, bench_promotion_lock, bench_generation_gc, bench_read_epoch, the 235-project resolver.
    • Not run locally: grammar-rebuild (no grammar changed; it needs CI's clang-18 toolchain).
  • Dogfood with the release binaries: find_callers(nearest_spellings) returns 4/4 resolved, including both lp:: calls in server.rs. The keyword phrase is matched literally.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj

Fixes what dogfooding v0.37.1 on this repository found, and the red nightly CI. The version is bumped in this PR with `version_bump.sh`. **Schema 81 → 82** (m0082: one full resolve on upgrade, no re-parse). Packages are unchanged. ## #367: Rust calls through a workspace module alias resolve (lane E) - **Cause:** the alias expansion feeding `temp.qual_groups` (`crates/indexer/src/index.rs`) only admitted modules spelled `.`/`crate::`/`self::`/`super::`. `use code_index_core::lang_profile as lp; lp::f()` was never expanded, although the fully spelled path resolved. - **Fix:** an aliased module is also admitted when its head passes `rust_use_head_is_workspace`, the check `follow_own_imports` already uses. The same-file join, the "alias declared twice binds nothing" gate and whole-qualifier matching are unchanged. The lookup reads Rust imports only, so it costs nothing in other languages (`fullscan_step` is identical to master). - **m0082** re-resolves once on upgrade. A schema-81 index of this repo migrates with `parsed=0` and equals a fresh index bind for bind. - **Measured, bind diff against master:** - rust-analyzer +50 `tier1q_alias`, all hand-checked correct. - this repo +68, 40/40 sample correct. - the other 8 pinned repos identical; nothing lost or retargeted. - `precision_gate` 7/7, phantom 0, with a new Rust alias probe and decoy. - **Re-records, all operator-approved:** - `tier3-baseline.json`: rust-analyzer only, exactly the +50. - `host-package-lifecycle.json`: schema condition only, 0 counts moved. - `ruby-package-cost.json`: schema 82. `vm_step` +0.006%, inside the record's 0.02% determinism. Measured on one box: master 27,933,085 / 27,932,224 against this binary 27,935,458 / 27,934,443. - `baseline.json` and `cost-baseline.json` unchanged. ## #368: `evidence_incomplete` only from partial files that could reference the target (lane P) - The single downgrade site (`downgrade_absence_verdict`) now counts a partial file only if its rows share the target's language (`filter_admits`), an active bridge leads into it, or its languages are unknown. - A package file that stopped before handing off its script regions counts for the host's route languages (TypeScript/JavaScript, from `EMBEDDED_ROUTES`). There is no path heuristic: a partial Rust file under `tests/` still downgrades a Rust target. - **Replies:** `partial_sources_in_scope` (a count) sits beside the unchanged `partial_sources_in_index`. The `resolution_gaps` census states its scope and gives `in_requested_scope` when the call is filtered. The agent-task token cost is exactly master's (12,495). ## Replies that state what they did not measure (lane D) - **#366:** `defining_file_verified` is set where the defining file is actually verified, for both `check_rename` and `safe_delete`. A full window gets a by-path pass in both tools. - **`changed_symbols` / `review_diff`:** a page with files remaining ships `omitted_symbols_unmeasured: "paged_files_not_counted"` instead of an unmeasured `0`. - **`spellings_present`:** when it is capped, it says "the N nearest of the M". - **`search_text`:** a bare `AND/OR/NOT/NEAR` is an operator only when single terms strictly alternate. `CREATE TABLE IF NOT EXISTS t` was silently answered as a boolean (339 files); it is now one literal phrase (1). `(a OR b) c` still works. **Behaviour change:** an unparenthesised `alpha OR beta gamma` is now a literal phrase. - **Recovery log:** "cannot be recovered" names what moved (packages, bytes, grants), or says the record cannot tell (host or config). ## Nightly CI (red since about 2026-10-06; fixed here) - **`bench_generation_gc`** (lane G): the bench promoted a hand-written, empty generation. 0.37.1's correct `Switch::owed` mark then re-extracted the files and emptied it, and the vacuity guard fired. The bench now builds the mirror generation for real. The guard is just as strict, and it still times 502/12,502 and 100,002/2,500,002 rows. - **`language_package_overhead`** (lane N): it compared native output against a binary pinned at 312a242e (2026-10-03). Every difference maps to an intended, blessed change: #309, #326, #335, #336, #337, #348, and the migration span fixes. Nothing was unexplained. - The check now grades the shipped no-package CLI against the in-process builtins. - Release-to-release drift stays owned by `corpus_ratchet`/`corpus_stage`; dropping a CLI builtin turns this check red. - The CI job no longer builds a second tree (−5.5 min). ## Verification (I re-ran each lane's tests and at least one mutation myself; all RED, restored by md5) - **One full local run**, plus every nightly-only CI step locally on the final tree 9a89ac6a: - fmt; clippy (workspace, first-party, windows-gnu, guest); rustdoc `--document-private-items -D warnings`; - workspace nextest 5137/5137, daemon leg 942/942, ci-timing 34/34; - the CI corpus job (11 suites), `psr4_vs_builtin`, tier 3, `corpus_scale`, mixed-load, `overview_scale_2m`; - `precision_gate` 7/7, phantom 0; - all 11 `package_leg.sh <pkg> full` and all 7 `overhead` legs. The js leg was re-measured quiet after a contended run at load 10. - The full `plugin-path-cost` job: phase2 acceptance (7 languages, signed install, MCP, watch, restart), `language_package_overhead`, `plugin_path_cost`, `package_pool`, `grammar_ab`, `claim_domain_scale`, `bench_promotion_lock`, `bench_generation_gc`, `bench_read_epoch`, the 235-project resolver. - Not run locally: `grammar-rebuild` (no grammar changed; it needs CI's clang-18 toolchain). - **Dogfood with the release binaries:** `find_callers(nearest_spellings)` returns 4/4 resolved, including both `lp::` calls in `server.rs`. The keyword phrase is matched literally. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
check_rename never set text_candidate_window.defining_file_verified although
text_occurrences verified the defining file inside the window, and had no
out-of-window pass when the window was saturated. The flag is now set inside
text_occurrences where the defining file's lines are checked, and the by-PATH
pass that used to live only in safe_delete moved into text_occurrences, so a
saturated window verifies the defining file for both tools with the same
per-line test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
The lang probe's spellings_present is cut at LANG_SPELLINGS_SHOWN (12), yet
the note said "every spelling held is in the lang probe's
spellings_present". The probe now carries spellings_present_total (counted
before the cap) and spellings_present_truncated, and both note arms name the
N nearest of the total when the list was cut.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
A page with later pages remaining shipped omitted_symbols: 0 while its own
hint said the remaining files' symbols are counted nowhere. Only a measured
cut is now counted: on the last page the count ships alone (0 earned); while
pages remain omitted_symbols_unmeasured: "paged_files_not_counted" ships,
with a count beside it only when this page's per-file cap really cut rows
(a floor). review_diff's coverage block applies the same values instead of
unwrap_or(0).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
safe_delete / check_rename downgraded every verdict to `evidence_incomplete`
whenever ANY file in the index was a partial source, so six intentionally
broken Svelte/XAML package fixtures made every Rust verdict in this repo
read "evidence incomplete".

One shared rule now decides (`partial_source_could_reference`, counted by
`partial_sources_relevant_to`): a partial file is in scope of a target when
its rows share the target's language (spelling-equivalent via the shared
lang_profile filter rule, so `ruby` == `de.h-dv.ruby/ruby`), when an active
granted bridge leads from one of its languages into the target's, or when
its languages / the bridges / the target are unknown (fail closed). No path
or test-role heuristic: a partial Rust file under tests/ still downgrades.

- daemon: `PartialSource.languages` = files.lang + extraction_components.lang
  of every live contribution (a Svelte script region counts as typescript);
  `PartialSources.bridges` read through bridges::load_active. Both additive,
  `None` from an older daemon = unknown = in scope.
- mcp: `evidence_gaps.partial_sources_in_scope {count, target_languages}` on
  safe_delete/check_rename beside the unchanged project-wide
  `partial_sources_in_index`, plus a semantics clause; the downgrade fires
  on the scoped count.
- resolution_gaps: `partial_source_census` keeps listing everything, gains
  a `scope` statement and `in_requested_scope` computed with the call's own
  project / exclude_tests (is_test_path, the rule that stamps test-role
  refs) / lang (filter_admits over the file's languages) / path_glob
  (SQLite GLOB).
- safe_delete/check_rename descriptions reworded shorter ("an in-scope
  file's extractor"), startup payload shrinks.

Tests: partial_source_scope_e2e (both legs) + unit tests; mutations run
and recorded in the test docs. harness_settle_barrier 58 -> 59.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
"none of N candidate activation(s) is the recorded one; something besides
[languages] moved" named nothing. recorded_activation now reads the
recorded generation's generation_packages rows and compares them with
today's packages: packages added/removed, container bytes, extraction
identity and capability grants are named per package key; when every
comparable input is equal, the line says the move is in the host/worker
identity or the project config, which the record keeps only inside the
digest. Still one log line and the NotRecovered Display.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Fail closed on the edge case fafd330e left open: a partial file whose
PACKAGE producer stopped early may have stopped before handing a
`<script>` region to the host, so the region's language is exactly what
the cut removed.

- indexer: `embedded::route_languages` resolves `EMBEDDED_ROUTES` through
  the compiled-in set the same way `delegate` does (typescript,
  javascript in this build) -- the host's whole delegation population.
- daemon: `PartialSource.unproven_region_languages` = those languages
  when a non-reserved package's WHOLE-FILE contribution (region_start 0)
  carries the diagnostics. Any guest may emit TAG_EMBEDDED_REGION (no
  manifest gate), so "package producer" is the recorded signal; a
  builtin's partial, or one only on a delegated region's own
  contribution (guest finished, regions proven), stays `None`.
- mcp: `partial_source_could_reference` unions them into the file's
  languages, so a broken .svelte downgrades a TypeScript target and,
  being bounded by the route table, still nothing in Rust. The census
  `lang` filter admits them too.

Tests: new e2e (both legs) on a lone Broken.svelte; unit test for the
bound; the existing TS expectation moves to count 3 with the reason.
Mutations run and recorded in the test docs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
`use code_index_core::lang_profile as lp;` then `lp::nearest_spellings()`
stayed unresolved although `imports` held {module, alias: "lp"}: the I051
alias expansion into `temp.qual_groups` admitted only modules whose
string starts with `.`/`crate::`/`self::`/`super::`, so a path headed by a
SIBLING workspace crate never entered it. The fully spelled path already
anchored (tier 1Q pass 1).

One rule, reusing clause (c) of `follow_own_imports`: an aliased module
whose head passes `rust_use_head_is_workspace` over the `file_anch`
package-tail vocabulary, in a language whose own-import proof is Rust's
module-file rule, is admitted too (`temp.alias_ws_modules`, read as one
indexed IN). Same-file join, ambiguity gate and whole-qualifier equality
are unchanged.

Measured bind for bind against the master binary, joined on (path, line,
col, kind, occurrence): rust-analyzer +50 (all tier1q_alias, all hand-
checked right), this repository +68 (40 sampled, seed 367, all right;
the seven lp:: calls in server.rs among them), the other eight pinned
repositories identical; 0 lost, 0 retargeted. corpus_ratchet and
corpus_cost green; tier3-baseline.json moves for rust-analyzer
(rule.tier1q_alias +50) and is NOT blessed here.

No migration: existing indexes gain these binds only when a Full
resolve runs. A re-heal needs a schema bump, which moves the schema
condition of the recorded package baselines; left to the operator.

Tests: resolution_provenance (positive, decoy pair, rebound, other-
file, external-crate negatives; mutations run), precision_gate Rust
probe + same-stem decoy (population 31/20/71/476 -> 33/21/74/483),
pool_capability_registry stance for the new temp relation.

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
Plain text holding an upper-case keyword between multi-word runs
("CREATE TABLE IF NOT EXISTS generation_packages") tripped
has_fts5_operator and was passed to FTS5 as an expression: loose words
around a NOT, a different and far broader query (339 files vs 1), with the
spanned word IF reported dark. A bare boolean is now an operator only when
single terms and operators strictly alternate (foo AND bar, a OR b OR c);
anything else is phrase-wrapped, as the quoted form already was.
whole_word_e2e's bare-expression case moves to the alternating form, and
its multi-word form becomes a literal control.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
MIGRATION 82 (`workspace_module_alias`), m0079's shape: no unbind, no
re-parse, one Full pass, skipped when no language's own imports are
proven by Rust's module-file rule. The resolver is NULL-filling, so
without it an upgraded index keeps `lp::nearest_spellings()` as
name_fallback until the file is edited. Measured: a master-written index
of this repository (schema 81) opened by this binary migrates to 82 with
parsed=0 and equals a fresh index bind for bind (91851 resolved both).
Test: `upgrading_gains_the_workspace_module_alias_binds` (an index left
in the older binary's state migrates to bind-for-bind equality with a
fresh one); MUTATION RUN: `up` returning before its Full pass -> RED.

COST: the first admission scanned `imports` and `file_anch` and always
spliced the IN arm; ruby_package_cost saw it on ruby-sinatra (no Rust
file): fullscan_step 351876 -> 352463. Now driven from `files` by the
Rust-proof languages `materialize_lang_profile` already computed (it
returns them), and the temp relation and the arm exist only when they
have a member: fullscan_step back to exactly 351876 package / 273464
builtin, vm_step inside master's own run-to-run spread. Corpus binds
re-diffed with this binary: rust-analyzer +50 tier1q_alias, the other
eight identical, 0 lost / 0 retargeted (unchanged from be1be13a).

RE-RECORDED, operator-approved:
- tests/corpus/host-package-lifecycle.json: condition only (schema
  81 -> 82), 0 categorised changes; COSI_PACKAGE_BLESS reason "#367
  schema 81->82, re-resolve on upgrade; condition only".
- tests/corpus/tier3-baseline.json: rust-analyzer only, exactly
  count.resolved +50, rule.tier1q_alias 47 -> 97, count.edges +40,
  count.resolved_type +43, count.resolved_call +6,
  count.resolved_import +1; py-django unchanged.

NOT re-recorded: tests/corpus/ruby-package-cost.json. Its trial bless
moved counted values (package vm_step 27930082 -> 27939870, inside the
documented ~0.02% determinism; wall_ratio_pct 131 -> 128 at load ~9),
so per the operator's rule it was reverted (md5 verified) and stays at
schema 81: ruby_package_cost is red on the schema condition until the
operator decides. baseline.json, cost-baseline.json, stage-baseline.json
untouched (md5 verified).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Condition only, operator-approved. fullscan_step is identical to schema
81; package_leg.vm_step 27930082 -> 27931686 (+0.006%), inside the
record's ~0.02% determinism. Measured on one box at load ~2: master
27933085 / 27932224, this binary 27935458 / 27934443. wall_ratio_pct
131 -> 135 (wall-clock; the previous bless moved it 133 -> 131).
baseline.json, cost-baseline.json, host-package-lifecycle.json and
tier3-baseline.json md5-unchanged by this commit.

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
Nightly run 1137 went RED in both arms: "generation 1 holds no
`file_contributions` rows". The mirror fixture hand-wrote a
reparse/owed queue for every file and promoted it without extracting
anything, so the incoming generation held nothing for any file it
claimed to re-extract. Since 6cfd094d the switch marks such a file
owed a pass (Switch::owed); the bench's control index_path then
re-upserted every files row and the files.id cascade took the
superseded generation's rows with it (owed=502, doomed 502/5000/6500/500
-> 0/0/0/0 measured at n=500). The product is right; the fixture was a
state no build can promote.

The mirror is now a real total rebuild (build::build_full under the
active builtin identity), so the incoming generation holds its own
rows, owed reads 0 and the collected generation keeps exactly the
rows it had before (same n+2 contributions and cascade). A new
`switch.owed == 0` assert names the cause at the switch; the guard at
the clock is unchanged. Setup cost rises by one extraction of the
tree, outside every clock. The active generation now holds a copy of
every module, which collect.rs's idle table predates; said there.

Mutation (run): restore the hand-written queue -> RED at the owed
assert (502); also neuter that -> RED at the clock guard, the nightly
message verbatim.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
The nightly `plugin-path-cost` job built a CLI from the FIXED commit
312a242e (the pre-migration release that qualified the runtime-to-package
extraction, 5d875ea0) and required today's native output to equal it,
admitting only four hand-written audits written for the migration's own
corrections. Every operator-approved native change since was blessed in
tests/corpus/baseline.json and had no audit here, so the test has been red
on every nightly since the first of them (run 1137 on the last release's
merge: rust-ripgrep, python-flask, ts-zod, ruby-sinatra).

Attribution, measured by indexing all seven tier-one repos with release
CLIs built at 312a242e, v0.35.0, v0.36.1 and this tree, row-diffed with the
corpus projection SQL joined per site:

  pin -> v0.35.0
    php-guzzle   2866 call span moves      migration terminal-span correction (5d875ea0, audited)
    rust-ripgrep  119 call span moves      migration terminal-span correction (anyhow::bail!)
                  410 unbind, 44 retarget  #336 (-410, 44 -> tier1r_receiver)
    python-flask  139 binding rows added   #309 module-scope bindings (71d10926)
                  348 unbind               #336 (-344) + lane P receiver evidence (-4)
    ts-zod         12 read rows added      #309 typeof property reads (71d10926)
                  131 unbind               #336 (-131)
    ruby-sinatra +1540 symbols, +859 refs  #335 RSpec outline + constant arguments
                  78 rule-111 binds, 24 rule-110 type binds   #326
                  185 unbind, 5 -> tier1r  #336
  v0.35.0 -> v0.36.1
    rust-ripgrep  228 binding qualifiers, 211 binding relocations,
                  147 rule-60 binds        #348 initializer type evidence (+147)
    ruby-sinatra  375 type -> member_access namespace heads, 47 rule-110
                  retargets, 48 method_call -> read           #337 (+423)
  v0.36.1 -> this tree: ZERO row differences on all seven repos.

Every class reconciles with a blessed record; nothing is unexplained.

The test now indexes each repo with this tree's release `code-index index`
(empty operator store) and in process with every builtin and no host,
and requires all five parity axes to be exact: the shipped CLI path is
the one thing the corpus suites never run. Release-to-release native
drift stays with corpus_ratchet and corpus_stage (per-push `corpus` job,
same seven repos), where it is already declared and blessed, so an
intended change needs no second declaration. The four migration audits
and their four unit tests go with the reference they graded.

Mutations, both run: filtering the ruby builtin out of `cmd_index`'s
plugins reds this test (ruby-sinatra 0 vs 2800 symbols, rust-ripgrep
4872 vs 4874) while corpus_ratchet cannot see it; emptying the Ruby
walker's emit_const_type_refs leaves this test green and reds
corpus_ratchet (ruby-sinatra refs -2091, resolved_type 429 -> 0).

CI: the step builds this tree's CLI into the job's own target/ instead
of fetching and cold-building a second tree; env CODE_INDEX_NATIVE_BASELINE
is replaced by CODE_INDEX_NATIVE_CLI, and the test refuses a CLI whose
version line is not this tree's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
a44ba75f cost plugin-wpf +136 tool tokens (12631 vs master 12495), into
the #235 reserve. Attributed and trimmed; the measurement stays:

- census rows no longer ship `languages` / `unproven_region_languages`:
  they are the scope rule's INPUTS, not its result (now `#[serde(skip)]`,
  graded on the daemon's own read instead -- new
  `a_partial_files_languages_come_from_its_contributions`).
- `partial_source_census.scope` + `in_requested_scope` appear only when
  the call actually passed `lang` / `path_glob` / `exclude_tests`; the
  scope statement is one short line.
- `evidence_gaps.partial_sources_in_scope` is a bare count (was
  `{count, target_languages}`); its explanatory sentence is replaced by a
  short one that fires only when the count is 0 and so spared the
  verdict.

plugin-wpf tool_tokens: 12631 -> 12495 (master 12495). Tests updated to
the trimmed shape; every mutation re-run RED (recorded in the test docs).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YDDBdWU4yThi8Zi51E6FCj
Merge lane P token trim (#368 reply shape back to master's cost) into train-0372
All checks were successful
Package CI / package-corpus (python) (pull_request) Successful in 28s
Package CI / package (python) (pull_request) Successful in 36s
Package CI / package-corpus (php) (pull_request) Successful in 36s
CI / cargo fmt (pull_request) Successful in 1m18s
Package CI / package (typescript) (pull_request) Successful in 33s
Package CI / package (timeline) (pull_request) Successful in 42s
Package CI / package (csharp) (pull_request) Successful in 45s
CI / OSS corpus tier-3 scale (nightly) (pull_request) Has been skipped
CI / Grammar rebuild from source (nightly) (pull_request) Has been skipped
Package CI / package-corpus (timeline) (pull_request) Successful in 37s
Package CI / package-corpus (csharp) (pull_request) Successful in 46s
Package CI / package-corpus (svelte) (pull_request) Successful in 46s
Package CI / package-corpus (typescript) (pull_request) Successful in 46s
Package CI / package-corpus (xaml) (pull_request) Successful in 48s
Package CI / package overhead budgets (nightly, one package at a time) (pull_request) Has been skipped
Package CI / package-corpus (ruby) (pull_request) Successful in 55s
Package CI / package-corpus (vue) (pull_request) Successful in 50s
CI / CI lane wall-clock headroom (pull_request) Successful in 1m29s
CI / guest crates (fmt, clippy, doc) (pull_request) Successful in 1m37s
CI / cargo doc (intra-doc links) (pull_request) Successful in 5m32s
CI / cargo check (MSRV 1.98) (pull_request) Successful in 6m51s
CI / cargo test (abi, 32-bit + wasm32) (pull_request) Successful in 7m0s
CI / cargo deny (pull_request) Successful in 7m1s
CI / cargo clippy (pull_request) Successful in 7m6s
CI / cargo check (windows-gnu) (pull_request) Successful in 7m25s
CI / cargo test (pull_request) Successful in 21m51s
CI / OSS corpus (tier 1) (pull_request) Successful in 31m55s
CI (Windows) / fmt + clippy + build + test (windows) (pull_request) Successful in 40m37s
CI / cargo test (daemon transport) (pull_request) Successful in 12m59s
CI / Plugin path cost + pool throughput (nightly) (pull_request) Has been skipped
9a89ac6aec
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!373
No description provided.