h-dv/code-index

Watch

Fork 0

#134's origin gate compares a kebab-case directory against a snake_case import, so hyphenated crates lose 58% of their cross-crate binds #165

New issue

Closed

opened 2026-09-05 23:56:52 +02:00 by buildagent · 2 comments

buildagent commented

2026-09-05 23:56:52 +02:00

Member

Found by refusing to bless the tier-3 corpus baseline on a mechanism argument and diffing the actual binds instead. This is a defect in #134 as merged (branch integration, commit 2f16e226), caught before release.

The comparison

crates/indexer/src/index.rs, in the import_key_rel build:

let tail_ok = pkg_tail_of.get(&t).is_some_and(|tail| {
    dotted
        .match_indices('.')
        .any(|(i, _)| dotted[..i].ends_with(tail.as_str()))
});

dotted is the dot-wrapped import module (.hir_expand.). tail is the package directory tail (hir-expand). "hir_expand".ends_with("hir-expand") is false, so pkg_tail_ok = 0.

pkg_tail_ok = 1 is the only disjunct of the tier-3 origin gate that admits a cross-package bind:

AND (fp.pkg_tail = ''
     OR COALESCE(fpr.pkg_dir, '') = fp.pkg_dir
     OR ikr.pkg_tail_ok = 1)

So every legitimate cross-crate reference into a hyphenated crate is gated off. Cargo derives the code identifier from the package name by folding - to _; nothing here does.

Measured on rust-analyzer (two full indexes, same corpus, base vs integration)

Base = 6f75e58 (pre-#125/#134). Both indexes reproduce the ratchet exactly (refs_resolved 127555 → 127867).

Resolved type binds: 52751 → 51288.

	count
LOST (in base, absent in integration)	2221
GAINED	758
of the LOST, same-crate	0 (0.0%)
of the LOST, cross-crate	2221 (100.0%)

Split by whether the target crate directory is hyphenated:

target crate dir	base	integration	delta
hyphenated (`hir-expand`, `hir-def`, …)	3216	1337	−1879 (−58%)
not hyphenated (`syntax`, `ide`, …)	8500	8218	−282 (−3.3%)

A 17× disparity, which is the signature.

A lost bind read at source — it is correct, not a phantom

crates/hir-def/src/attrs/docs.rs:331 — file_id: HirFileId lost its bind to crates/hir-expand/src/lib.rs.

use hir_expand::{
    AstId, ExpandTo, HirFileId, InFile,
    ...
    span_map::SpanMap,
};

The file explicitly imports the symbol, and hir-def/Cargo.toml declares hir-expand.workspace = true. The import vouches for the origin exactly as the gate intends; the gate cannot see it because of the spelling.

The gate is right about other things in the same file. Crate at docs.rs:327 lost a bind to crates/hir/src/lib.rs and that removal is correct — Crate is imported from base_db (line 14), and hir-def does not depend on hir at all (that edge runs the other way). So #134 is doing real work; this defect is riding along with it.

Why no existing gate caught it

The tier-1 corpus structurally cannot see it. ripgrep's crates are core, globset, grep, ignore, matcher, pcre2, printer, regex, searcher, cli — not one is hyphenated. That is why the lane's bind inspection of ripgrep found only 2 correct binds lost and reasonably concluded the change was a precision gain. It was, on that corpus. rust-analyzer is the first pinned repo with kebab-case crate directories, and it is in the tier-3 suite, which is why the tier-3 ratchet is what went red.

This repo is affected too: crates/mcp-server, crates/plugin-host, crates/plugin-supervisor are all hyphenated, so cosi-mcp degrades its own resolution.

Suggested fix — one rule, both ends

Fold - to _ on both sides of the comparison, at the two points where the two vocabularies meet: when pkg_tail is stored into temp.file_pkg, and in the segment-wrapping that produces IMPORT_SEGMENTS/QUALIFIER_SEGMENTS. Then every comparison site (tier 1b's file-key arm, tier 1Q, tier 3 Edge 2, tier 1R) inherits it without its own patch — this should be one structural clause, not a fix per tier.

It is not Rust-only, which is part of why it belongs at the vocabulary boundary rather than in a language arm: Python distributions are routinely python-dateutil → dateutil, scikit-learn → sklearn; Ruby has activesupport → active_support.

Bound the claim honestly: folding can only add matches, so it cannot remove a bind — but it can admit one, and a foo-bar/foo_bar pair in one tree would collide. State that in the code rather than assuming it away.

Required before this is called fixed

Re-measure both baselines and inspect the new binds at source — the numbers below are pre-fix and will move.
The tier-1 baseline.json re-record on integration (8e15f983…) was blessed against the pre-fix delta and must be redone, not amended in place.
Add a hyphenated-crate fixture to the precision gate, so this cannot regress invisibly again. The absence of one is the reason a 58% recall loss reached a bless decision.

🤖 Generated with Claude Code

https://claude.ai/code/session_01K1zj5VcFJvJt3pQxe9259K

Found by refusing to bless the tier-3 corpus baseline on a mechanism argument and diffing the actual binds instead. **This is a defect in #134 as merged** (branch `integration`, commit `2f16e226`), caught before release. ## The comparison `crates/indexer/src/index.rs`, in the `import_key_rel` build: ```rust let tail_ok = pkg_tail_of.get(&t).is_some_and(|tail| { dotted .match_indices('.') .any(|(i, _)| dotted[..i].ends_with(tail.as_str())) }); ``` `dotted` is the dot-wrapped **import module** (`.hir_expand.`). `tail` is the **package directory tail** (`hir-expand`). `"hir_expand".ends_with("hir-expand")` is **false**, so `pkg_tail_ok = 0`. `pkg_tail_ok = 1` is the only disjunct of the tier-3 origin gate that admits a *cross-package* bind: ```sql AND (fp.pkg_tail = '' OR COALESCE(fpr.pkg_dir, '') = fp.pkg_dir OR ikr.pkg_tail_ok = 1) ``` So every legitimate cross-crate reference into a **hyphenated** crate is gated off. Cargo derives the code identifier from the package name by folding `-` to `_`; nothing here does. ## Measured on rust-analyzer (two full indexes, same corpus, base vs integration) Base = `6f75e58` (pre-#125/#134). Both indexes reproduce the ratchet exactly (`refs_resolved` 127555 → 127867). Resolved **type** binds: 52751 → 51288. | | count | |---|---| | LOST (in base, absent in integration) | **2221** | | GAINED | 758 | | of the LOST, **same-crate** | **0 (0.0%)** | | of the LOST, **cross-crate** | **2221 (100.0%)** | Split by whether the *target crate directory* is hyphenated: | target crate dir | base | integration | delta | |---|---|---|---| | **hyphenated** (`hir-expand`, `hir-def`, …) | 3216 | 1337 | **−1879 (−58%)** | | not hyphenated (`syntax`, `ide`, …) | 8500 | 8218 | −282 (−3.3%) | A **17× disparity**, which is the signature. ## A lost bind read at source — it is correct, not a phantom `crates/hir-def/src/attrs/docs.rs:331` — `file_id: HirFileId` lost its bind to `crates/hir-expand/src/lib.rs`. ```rust use hir_expand::{ AstId, ExpandTo, HirFileId, InFile, ... span_map::SpanMap, }; ``` The file **explicitly imports the symbol**, and `hir-def/Cargo.toml` declares `hir-expand.workspace = true`. The import vouches for the origin exactly as the gate intends; the gate cannot see it because of the spelling. **The gate is right about other things in the same file.** `Crate` at `docs.rs:327` lost a bind to `crates/hir/src/lib.rs` and that removal is **correct** — `Crate` is imported from `base_db` (line 14), and hir-def does not depend on `hir` at all (that edge runs the other way). So #134 is doing real work; this defect is riding along with it. ## Why no existing gate caught it **The tier-1 corpus structurally cannot see it.** ripgrep's crates are `core`, `globset`, `grep`, `ignore`, `matcher`, `pcre2`, `printer`, `regex`, `searcher`, `cli` — **not one is hyphenated**. That is why the lane's bind inspection of ripgrep found only 2 correct binds lost and reasonably concluded the change was a precision gain. It was, *on that corpus*. rust-analyzer is the first pinned repo with kebab-case crate directories, and it is in the **tier-3** suite, which is why the tier-3 ratchet is what went red. **This repo is affected too**: `crates/mcp-server`, `crates/plugin-host`, `crates/plugin-supervisor` are all hyphenated, so cosi-mcp degrades its own resolution. ## Suggested fix — one rule, both ends Fold `-` to `_` on **both** sides of the comparison, at the two points where the two vocabularies meet: when `pkg_tail` is stored into `temp.file_pkg`, and in the segment-wrapping that produces `IMPORT_SEGMENTS`/`QUALIFIER_SEGMENTS`. Then every comparison site (tier 1b's file-key arm, tier 1Q, tier 3 Edge 2, tier 1R) inherits it without its own patch — this should be one structural clause, not a fix per tier. It is not Rust-only, which is part of why it belongs at the vocabulary boundary rather than in a language arm: Python distributions are routinely `python-dateutil` → `dateutil`, `scikit-learn` → `sklearn`; Ruby has `activesupport` → `active_support`. **Bound the claim honestly**: folding can only *add* matches, so it cannot remove a bind — but it can admit one, and a `foo-bar`/`foo_bar` pair in one tree would collide. State that in the code rather than assuming it away. ## Required before this is called fixed - Re-measure **both** baselines and inspect the new binds **at source** — the numbers below are pre-fix and will move. - The tier-1 `baseline.json` re-record on `integration` (`8e15f983…`) was blessed against the pre-fix delta and **must be redone**, not amended in place. - Add a **hyphenated-crate fixture** to the precision gate, so this cannot regress invisibly again. The absence of one is the reason a 58% recall loss reached a bless decision. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01K1zj5VcFJvJt3pQxe9259K

rule	pre-#134 `6f75e58`	integration `348924a`	fixed	fix − pre-#134
`tier1b_file_key_import`	15308	15308	23172	+7864
`tier1b_name_import`	5084	5084	7190	+2106
`tier1b_container_import`	1490	1490	1640	+150
`tier1q_pass1`	35335	36314	39843	+4508
`tier3_import_boost`	10321	9617	11270	+949
`tier1q_tiebreak`	363	363	427	+64
`tier1r_receiver`	1951	1922	1802	−149
total	127555	127867	143113	+15558

artifact	md5
`tests/corpus/baseline.json`	`534084b8…` → `914dda1aa1eea933f72a20d2287031f0`
`tests/corpus/tier3-baseline.json`	`210e7e1f…` → `94abf592dabe654166b38c46a845fd11`
`tests/corpus/stage-baseline.json`	`a3e90a1e…` → `7d9695befe2e5968ac4ab05d403104ac`

Rows
Columns

#134's origin gate compares a kebab-case directory against a snake_case import, so hyphenated crates lose 58% of their cross-crate binds #165

The comparison

Measured on rust-analyzer (two full indexes, same corpus, base vs integration)

A lost bind read at source — it is correct, not a phantom

Why no existing gate caught it

Suggested fix — one rule, both ends

Required before this is called fixed

This is not a defect in #134. It is 47 days and 64 releases older.

The mechanism, and why it is one clause

Measured, on rust-analyzer, read at source

Phantom bound: 2 of 16 717 (0.012 %), found by an oracle outside the resolver

The gate that was structurally blind now is not

Baselines

Gates

Closing — fixed, merged, and verified independently of the lane's report

The correction to this issue's own title and body, for the record

Why the corpus could not see it

Residual, named not hidden