feat(parser): add Rust support

[tuannx]

Summary

• add a tree-sitter Rust parser for structs, enums, unions, traits, type aliases, functions, methods, imports, qualified references, trait inheritance, and implementations
• model Rust module conventions and Cargo workspaces, including member-crate prefixes and single-crate src roots
• defer impl-owner resolution so sibling-module aliases resolve correctly while external and blanket impls do not manufacture local entities
• wire Rust through ingest/autodetection, caching, CLI/MCP/action help, optional language dependencies, and docs

Addresses the Rust parser roadmap item tracked in #13. The status-board issue should remain open.

Accuracy review

The regression suite covers raw identifiers, unions, async/unsafe/extern functions, inline modules, super glob imports, aliases, generic owners, cfg-alternative impls, external qualified owners, blanket generic impls, Cargo workspaces, source-root detection, and cache invalidation.

Two false-positive cases found during review are explicitly guarded:

• std::io::Error cannot resolve to an unrelated local Error
• impl<T> Trait for &mut T cannot attach methods to an unrelated local struct T

Known static-analysis boundaries: macro-generated items are not expanded, cfg alternatives are structurally merged, and external dependency entities are intentionally not added.

Validation

• .venv/bin/python -m pytest -q — 261 passed
• .venv/bin/python -m ruff check src/ tests/ — passed
• focused mypy review — no errors in parsers/rust.py (remaining findings are pre-existing in five imported modules)
• fresh end-to-end analysis on BurntSushi/ripgrep at 227381db0ee83dfa4341f1e27ff9617c0f5ad992 (66k+ stars):
  • 86 production Rust files
  • 0 tree-sitter ERROR nodes and 0 missing nodes
  • 3,074 entities, 3,487 deduplicated edges, 12 recovered components
  • 0 missing method owners, duplicate package memberships, or dangling edges
• wheel build verified arcade_agent/parsers/rust.py is packaged

[github-actions]

Architecture Drift Report

Algorithm: PKG | Entities: 719 | Components: 7

Drift from Baseline

Metric	Baseline	Current	Delta
Components	7	7	+0
Similarity	—	0.98	—
BalancedArchitectureScore	—	0.68	—
PrincipleAlignmentScore	—	0.84	—
RCI	—	0.99	—
TurboMQ	—	0.24	—
BasicMQ	—	0.24	—
IntraConnectivity	—	0.00	—
InterConnectivity	—	0.01	—
TwoWayPairRatio	—	0.00	—
DependencyHealth	—	0.99	—
ComponentBalance	—	0.22	—
HubBalance	—	0.67	—
BoundaryClarity	—	1.00	—
DependencyDistribution	—	0.57	—
SmellDiscipline	—	0.91	—

Changes

• 49 entity movement(s) between components

Smells (1)

• Concern Overload: Tests

---

Generated by arcade-agent

[github-actions]

🤖 Architecture Analysis Summary

Powered by arcade-agent — automatic architectural self-analysis

---

📈 Metric Evolution

Baseline commit: 613a1d9

Legend: 🟢 better · 🔴 worse · 🟡 low impact · ⚪ no change

Metric	Baseline	Current	Change
BalancedArchitectureScore	0.6193	0.6848	🟢 ↑ (+0.0655)
📦 Components	9	9	⚪ → (no change)
🧩 Entities	115	120	🟡 ↑ (+5.0000)
🔗 Edges	115	120	🟡 ↑ (+5.0000)
🏷️ Classes	22	27	🟡 ↑ (+5.0000)
ƒ Functions	271	285	🟡 ↑ (+14.0000)
🔧 Methods	34	39	🟡 ↑ (+5.0000)
RCI	0.5652	0.7417	🟢 ↑ (+0.1765)
TurboMQ	0.2493	0.3174	🟢 ↑ (+0.0681)
BasicMQ	0.2493	0.3174	🟢 ↑ (+0.0681)
IntraConnectivity	0.0182	0.0168	🟡 ↓ (-0.0014)
InterConnectivity	0.0323	0.0340	🔴 ↑ (+0.0017)
TwoWayPairRatio	0.0000	0.0000	⚪ → (no change)
DependencyHealth	0.9822	0.9813	🔴 ↓ (-0.0009)
ComponentBalance	0.5304	0.5204	🔴 ↓ (-0.0100)
HubBalance	0.5000	0.5000	⚪ → (no change)
BoundaryClarity	0.9444	0.9444	⚪ → (no change)
DependencyDistribution	0.6263	0.6049	🔴 ↓ (-0.0214)
SmellDiscipline	0.9333	1.0000	🟢 ↑ (+0.0667)
PrincipleAlignmentScore	0.8553	0.8625	🟢 ↑ (+0.0072)

🏛️ Current Architecture

Metric	Value
📦 Components	9
🧩 Entities	120
🔗 Edges	120
🏷️ Classes	27
ƒ Functions	285
🔧 Methods	39
Balanced Score 🟡	0.6848 (Fair)
Principle Alignment	0.8625
RCI	0.7417
TurboMQ	0.3174
BasicMQ	0.3174
IntraConnectivity	0.0168
InterConnectivity	0.0340
TwoWayPairRatio	0.0000
DependencyHealth	0.9813
ComponentBalance	0.5204
HubBalance	0.5000
BoundaryClarity	0.9444
DependencyDistribution	0.6049
SmellDiscipline	1.0000

🧭 Principle Signals

Signal	Value
AcyclicDependencies	1.0000
LayeringHealth	0.9860
ResponsibilityFocus	1.0000
InterfaceSegregation	1.0000
ComponentBalance	0.5204
HubBalance	0.5000
BoundaryClarity	0.9444
DependencyDistribution	0.6049
SmellDiscipline	1.0000

🎯 Score Drivers

Biggest risks

• HubBalance: gap=0.5000 (signal=0.5000)
• ComponentBalance: gap=0.4796 (signal=0.5204)
• DependencyDistribution: gap=0.3951 (signal=0.6049)

Strongest areas

• AcyclicDependencies: gap=0.0000 (signal=1.0000)
• InterfaceSegregation: gap=0.0000 (signal=1.0000)
• ResponsibilityFocus: gap=0.0000 (signal=1.0000)

🕸️ High-Level Design

graph TD
    Algorithms["Algorithms\n43 entities\n5 classes / 3 methods"]
    Budget["Budget\n3 entities\n0 classes / 0 methods"]
    Cache["Cache\n4 entities\n0 classes / 0 methods"]
    Ci["Ci\n5 entities\n0 classes / 0 methods"]
    Exporters["Exporters\n15 entities\n1 classes / 0 methods"]
    Incremental["Incremental\n2 entities\n1 classes / 2 methods"]
    Parsers["Parsers\n22 entities\n16 classes / 32 methods"]
    Serialization["Serialization\n8 entities\n0 classes / 0 methods"]
    Tools["Tools\n18 entities\n4 classes / 2 methods"]
    Cache --> Serialization
    Ci --> Algorithms
    Ci --> Exporters
    Ci --> Serialization
    Ci --> Tools
    Serialization --> Algorithms
    Serialization --> Parsers
    Tools --> Cache
    Tools --> Parsers

Loading

🏗️ Components breakdown

Component	Entities	Classes	Methods
Algorithms	43	5	3
Parsers	22	16	32
Tools	18	4	2
Exporters	15	1	0
Serialization	8	0	0
Ci	5	0	0
Cache	4	0	0
Budget	3	0	0
Incremental	2	1	2

🚨 Architectural Smells

✅ No architectural smells detected.

📈 Evolution vs Baseline

Baseline commit: 613a1d9

Architecture-to-Architecture (A2A) Comparison

Metric	Value
A2A Similarity	0.9328
Matched Components	9
Components Added	0
Components Removed	0

Component matching details

Matched:

Baseline	Current	Similarity
Budget	Budget	1.0000
Cache	Cache	1.0000
Ci	Ci	1.0000
Incremental	Incremental	1.0000
Serialization	Serialization	1.0000
Exporters	Exporters	0.9333
Algorithms	Algorithms	0.9070
Tools	Tools	0.7826
Parsers	Parsers	0.7727

High-level component statistics

Status	Baseline	Current	Similarity	Entities	Classes	Methods
matched	Algorithms	Algorithms	0.9070	39 → 43 (+4)	5 → 5 (0)	3 → 3 (0)
matched	Budget	Budget	1.0000	3 → 3 (0)	0 → 0 (0)	0 → 0 (0)
matched	Cache	Cache	1.0000	4 → 4 (0)	0 → 0 (0)	0 → 0 (0)
matched	Ci	Ci	1.0000	5 → 5 (0)	0 → 0 (0)	0 → 0 (0)
matched	Exporters	Exporters	0.9333	14 → 15 (+1)	1 → 1 (0)	0 → 0 (0)
matched	Incremental	Incremental	1.0000	2 → 2 (0)	1 → 1 (0)	2 → 2 (0)
matched	Parsers	Parsers	0.7727	17 → 22 (+5)	11 → 16 (+5)	27 → 32 (+5)
matched	Serialization	Serialization	1.0000	8 → 8 (0)	0 → 0 (0)	0 → 0 (0)
matched	Tools	Tools	0.7826	23 → 18 (-5)	4 → 4 (0)	2 → 2 (0)

Before/After Mermaid diagrams

Baseline

graph TD
    Algorithms["Algorithms\n39 entities\n5 classes / 3 methods"]
    Budget["Budget\n3 entities\n0 classes / 0 methods"]
    Cache["Cache\n4 entities\n0 classes / 0 methods"]
    Ci["Ci\n5 entities\n0 classes / 0 methods"]
    Exporters["Exporters\n14 entities\n1 classes / 0 methods"]
    Incremental["Incremental\n2 entities\n1 classes / 2 methods"]
    Parsers["Parsers\n17 entities\n11 classes / 27 methods"]
    Serialization["Serialization\n8 entities\n0 classes / 0 methods"]
    Tools["Tools\n23 entities\n4 classes / 2 methods"]
    Cache --> Serialization
    Ci --> Algorithms
    Ci --> Exporters
    Ci --> Serialization
    Ci --> Tools
    Serialization --> Algorithms
    Serialization --> Parsers
    Tools --> Algorithms
    Tools --> Cache
    Tools --> Exporters
    Tools --> Parsers

Loading

Current

graph TD
    Algorithms["Algorithms\n43 entities\n5 classes / 3 methods"]
    Budget["Budget\n3 entities\n0 classes / 0 methods"]
    Cache["Cache\n4 entities\n0 classes / 0 methods"]
    Ci["Ci\n5 entities\n0 classes / 0 methods"]
    Exporters["Exporters\n15 entities\n1 classes / 0 methods"]
    Incremental["Incremental\n2 entities\n1 classes / 2 methods"]
    Parsers["Parsers\n22 entities\n16 classes / 32 methods"]
    Serialization["Serialization\n8 entities\n0 classes / 0 methods"]
    Tools["Tools\n18 entities\n4 classes / 2 methods"]
    Cache --> Serialization
    Ci --> Algorithms
    Ci --> Exporters
    Ci --> Serialization
    Ci --> Tools
    Serialization --> Algorithms
    Serialization --> Parsers
    Tools --> Cache
    Tools --> Parsers

Loading

Component dependency delta

Status	Source	Target
removed	Tools	Algorithms
removed	Tools	Exporters

Smell changes:

• ✅ Resolved: Concern Overload

💡 CI/CD Insights

• Quality Score: 🟡 Fair (BalancedArchitectureScore=0.6848)
• Principle Alignment: 0.8625 (higher means cleaner layering, focus, and boundaries)
• Top Risk Driver: HubBalance (signal=0.5000)
• Trend: 📈 Improving modularity
• Architecture Stability: 🟢 High (A2A=0.9328)
• Smells: ✅ Clean — no architectural smells

📄 View HTML reports and artifacts

---

This comment is auto-generated by the self-dogfooding CI job. It updates on every push to this PR.

[Copilot]

Pull request overview

Adds first-class Rust parsing support to arcade-agent, integrating a new tree-sitter-based Rust parser into ingest/parse flows, caching, and project documentation so Rust projects (including Cargo workspaces) can be analyzed like existing languages.

Changes:

• Introduces RustParser with module-convention handling, entity extraction (types/traits/functions/methods), and a second-pass linker for imports/references/trait relationships.
• Extends language detection/ingest, parse caching, and CI/MCP/action help text to include rust.
• Updates docs/roadmap and adds a Rust-focused regression test suite (including workspace + cache invalidation checks).

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/test_parsers/test_rust.py	Adds Rust parser regression coverage across module conventions, impl ownership, imports, and Cargo workspace behaviors.
tests/test_cache.py	Ensures cache key changes when Rust source files change.
src/arcade_agent/tools/ingest.py	Adds Rust file extensions and Cargo workspace-aware source-root detection.
src/arcade_agent/tools/adapters/mcp.py	Updates MCP adapter docstrings/help to list Rust as a supported language override.
src/arcade_agent/parsers/rust.py	New Rust tree-sitter parser + linker implementation with Cargo workspace/module modeling.
src/arcade_agent/parsers/init.py	Registers Rust parser behind an optional dependency import guard.
src/arcade_agent/ci/run_self_analysis.py	Updates CLI help to include Rust language override.
src/arcade_agent/cache.py	Includes .rs in cache-key file hashing.
ROADMAP.md	Marks the Rust parser roadmap item as shipped and describes scope.
README.md	Documents Rust support and updates installation extras to include [languages].
pyproject.toml	Adds tree-sitter-rust to the optional languages extra.
actions/analyze/action.yml	Extends action input docs to include Rust as a language override.
.github/workflows/architecture-analysis-reusable.yml	Extends reusable workflow input docs to include Rust as a language override.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[lemduc]

Strong parser overall — and I want to lead with what you got right: defect classes 2–4 from the Kotlin review (function-local hoisting, cross-package leaf-fallback edges, lost inheritance/shared props) are all clean here, with regression tests. resolve() returning None for unresolved multi-segment paths, the blanket-generic-impl skip, and the cfg-alternative dedupe are exactly the right instincts. Registration surfaces are complete, and the ripgrep validation numbers in the PR body held up under independent re-run.

One blocking finding, and it's an ironic one:

1. [Blocking] The recursion defect class from the Kotlin review is back — in four places. _references is iterative (so the original deep-parens killer passes — proof you know the pattern), but these four still recurse unboundedly, and extraction runs outside the per-file try (rust.py:475-481 wraps only stat/read/parse, catching only (OSError, ValueError)):

Crash input (one ~5KB poisoned file)	Site	Depth
a::a::…::T path, 991 segments	_path_segments (~L131)	991
use a::{a::{…}}, 991 levels	_flatten_use (~L160)	991
mod m { mod m { … }}, 993 levels	visit_container (~L456, mod_item)	993
impl T for &&&…&R / ((((R)))), 993 wrappers	_base_type_path (~L224)	993

Each reproduces RecursionError out of RustParser().parse(), killing the entire analysis and losing the good sibling files' entities. Fix pattern is already on main from the Kotlin follow-up (b7effc5): explicit stacks. Please also widen the per-file except to Exception like kotlin.py:374 as a backstop.

2. [Decide before merge] #[cfg(test)] mod tests contents become production entities. Rust unit tests live inline in production files, and exclude_tests=True only filters test paths — on a fresh ripgrep clone, 548 of 3,184 entities (17%) and 37 phantom packages (e.g. grep_cli.escape.tests.backslash) are test items, which will surface as bogus components in recovery and skew metrics. Suggested: when exclude_tests, skip mod_items whose preceding sibling attribute is #[cfg(test)] — or at minimum document the behavior. Maintainer call since it changes graph shape.

Non-blocking:

• _MAX_FILE_BYTES = 1_000_000 silently drops >1MB files from the graph — new behavior no other parser has, undocumented (and it doesn't mitigate the recursion crashes; the 991-segment file is ~5KB).
• README labels Rust "(full)" while Kotlin is "(structural)" — same tier, inconsistent label.
• rel_path at ~L483 duplicates the relative_to already computed inside the try.
• Per-file except tuple differs from kotlin.py's — align while fixing finding 1.

Also a heads-up on land order: #19 (polyglot) structurally rewrites ingest.py, and your touches there are small — expect to rebase this onto #19 rather than the reverse; while doing so, eyeball the interaction between your Cargo-workspace root detection in _detect_source_root and #19's multilang path (which deliberately bypasses it).

Test results on the branch: 290 passed, ruff clean; deep-parens/local-types/qualified-external-leaf/macros/invalid-file adversarial runs all clean except the four recursion inputs above. Fix finding 1, decide finding 2, and this is merge-ready.

[tuannx]

Addressed all requested changes in 55d9760 after rebasing onto main@8090d12.

Blocking fixes

• Replaced all four unbounded recursive walkers (_path_segments, _flatten_use, inline module traversal, and _base_type_path) with iterative stack/worklist implementations.
• Made extraction transactional per file: a malformed or adversarial file is logged and discarded without leaking partial graph state or losing healthy sibling files.
• Excluded inline #[cfg(test)] mod tests trees from the production graph while retaining non-test cfg modules.

Inline review fixes

• Switched Cargo manifest reads to binary tomllib.load() with safe malformed-encoding/TOML handling in both ingest and the Rust parser.
• Added Cargo.toml to Rust cache keys, including explicit .rs file-list analysis.
• Added regression coverage for both behaviors; the three outdated inline threads have been replied to and resolved.

Recurrence prevention and dogfooding

• Added depth-1,200 adversarial tests for all four recursive shapes, plus poisoned-file/healthy-sibling isolation tests.
• A depth-5,000 qualified-path case improved from ~11.3s to ~0.008s after removing recursive/repeated traversal.
• Re-ran against ripgrep (86 production Rust files): 2,641 entities / 2,215 edges, zero test-named entities/packages, missing owners, dangling edges, or duplicate package members.
• Ran Arcade Agent against itself before/after: architecture graph and metrics remained stable (120 entities / 120 edges; Balanced Architecture Score 0.6229).
• Added docs/BUG_CATALOG.md documenting the repeated Kotlin/Rust recursion defect and .github/skills/harden-tree-sitter-parsers/SKILL.md as the reusable parser-hardening workflow.

Validation

• 306 passed
• Ruff clean
• mypy clean for touched source files
• wheel contains the Rust parser
• Python 3.12/3.13 CI and Architecture Drift Detection are green

The PR is currently MERGEABLE / CLEAN and ready for re-review.

[tuannx]

Resolved the remaining Concern Overload: Tools smell in 86fb591.

Root cause: package recovery correctly recognized thin facades by outgoing dependency affinity, but any caller from another component prevented reassignment. This left compare, compute_metrics, detect_smells, recover, and visualize in the broad Tools bucket even though each fronts a single subsystem.

The refinement now treats incoming callers as usage, not responsibility, only for oversized package buckets. Compact boundaries remain package-anchored, so focused components such as Cache are not pulled into their implementation dependencies. Regression tests cover both the oversized-bucket reassignment and compact-boundary guard.

Self-dogfood delta:

Metric	Before	After
Architectural smells	1	0
Tools entities	23	18
RCI	0.5833	0.7417
TurboMQ	0.2504	0.3174
BalancedArchitectureScore	0.6229	0.6848
Entities / edges / components	120 / 120 / 9	120 / 120 / 9

Validation: 308 passed, Ruff clean, and git diff --check clean. The CI self-analysis should refresh the PR report from this commit.

[lemduc]

Review follow-up — per-finding status

Verified each finding against the branch as it stands (three commits landed after the review: ed9c878, 55d9760, 86fb591), then implemented what was left. New commit: cbea67f.

#	Finding	Status	Where
1	[Blocking] Unbounded recursion in _path_segments, _flatten_use, visit_container/mod_item, _base_type_path	Already fixed before this pass	55d9760
1b	Per-file except (OSError, ValueError) → Exception backstop	Already fixed	55d9760
2	[Decide before merge] #[cfg(test)] mod tests become production entities	Fixed now — and see the behavior note below	cbea67f
NB-1	_MAX_FILE_BYTES = 1_000_000 silent >1MB drop	Fixed now — removed	cbea67f
NB-2	README labels Rust "(full)" vs Kotlin "(structural)"	Already fixed (both now "structural")	55d9760
NB-3	rel_path duplicates the relative_to inside the try	Already fixed	55d9760
NB-4	Per-file except tuple differs from kotlin.py	Already fixed (both except Exception)	55d9760

Finding 1 — verified, not just assumed

All four sites are explicit LIFO worklists now, and extraction moved inside the per-file try. I re-ran the exact crash inputs from the review table against the current code:

Input	Result
a::a::…::T, 991 segments	parses, 1 entity
use a::{a::{…}}, 991 levels	parses, 1 entity
mod m { mod m { … }}, 993 levels	parses, 2 entities
impl T for &&&…&R, 993 wrappers	parses, 2 entities
impl T for ((((R)))), 993 wrappers	parses, 2 entities

No RecursionError. The existing parametrized regression test already covers the first four at depth 1,200; I added the parenthesized-type shape (((((R))))) as a fifth case since it was listed in the table but not covered.

⚠️ Finding 2 — behavior change, please confirm

55d9760 already skipped #[cfg(test)] modules, but unconditionally. That silently overrode exclude_tests=False, so I gated it:

• LanguageParser gains an exclude_tests attribute (default True), documented as "drop inline test constructs"; parsers for languages without inline tests ignore it.
• RustParser skips a mod_item only when it carries #[cfg(test)] and self.exclude_tests. The whole subtree goes — nested mods and impl blocks inside it too.
• The flag is threaded analyze → parse tool → parser, and folded into the parse cache key so the two graph shapes can't collide in .arcade-cache.

This changes graph shape on Rust repos — on ripgrep it removes ~548 of 3,184 entities (17%) and ~37 phantom packages from default runs. That is the intent (they are test items, and Rust unit tests are inline so path-based exclusion can never see them), but it is a visible default-behavior change and you can veto it: reverting is a one-line change to the gate. exclude_tests=False now restores the old output exactly.

Non-blocking NB-1 — chose removal over documentation

Removed _MAX_FILE_BYTES. Reasons: no other parser caps input size, it drops real code silently with only a log line, and it never mitigated the recursion crashes it looked like it was guarding — the per-file except Exception boundary is the actual backstop. The README had documented the cap; that text is corrected. Added a design-checklist bullet to docs/BUG_CATALOG.md so the pattern isn't reintroduced. Regression test asserts a >1MB file still yields its entities.

Verification

• pytest: 314 passed (baseline before this work: 308 passed — +6 new tests)
• ruff check src/ tests/: All checks passed
• mypy src/: 220 errors before and after, byte-identical set in the touched files — pre-existing repo-wide, no regression from this change

Deliberately left undone

• No rebase/merge with feat: polyglot multilang parse (roadmap #18 MVP) #19 — land order is a merge-time concern.
• .arcade/baseline.json not regenerated; that is the arch-drift workflow's job on main, and the branch has not touched it.

[tuannx]

Posting a self-review before merge — I verified the branch in an isolated worktree: 314 tests pass, ruff clean, mypy clean on parsers/rust.py. Three things worth resolving first, and I'd like your call on two of them, @lemduc.

1. Confirmed bug: comment between #[cfg(test)] and its module defeats test exclusion

In visit_container, pending_attributes is cleared by any non-attribute node, and tree-sitter emits comments as named children. Reproduced against this branch:

pub struct Production;
#[cfg(test)]
// unit tests for this module
mod tests { struct Fixture; }

→ tests.Fixture leaks into the graph despite exclude_tests=True. Stacked attributes (#[cfg(test)] + #[allow(...)]) work correctly; only interleaved comments break it.

Fix is one line — skip line_comment/block_comment nodes without clearing pending_attributes — plus a regression test in the adversarial matrix. I'll push this to the branch.

2. recover.py facade-refinement change is out of scope for this PR — your call

The change (external callers no longer pin an entity to an oversized bucket, sharing CONCERN_OVERLOAD_ENTITY_THRESHOLD from concern.py) alters pkg-algorithm recovery output for every language, not just Rust, and the PR description doesn't mention it. It has tests covering both directions and is likely what made the ripgrep components settle at 12 — but it silently shifts baselines for existing users.

Options:

• a) keep it here but document it prominently in the PR body / changelog, or
• b) I split it into its own PR so the behavior change gets its own review trail.

I lean (b) for changelog honesty, but happy either way — @lemduc which do you prefer?

3. Release sequencing: docs advertise Rust against 0.2.0, which doesn't contain it

The README now lists Rust support and bumps the pins to actions/analyze@v0.2.0 / arcade-agent-version: "0.2.0" — but released 0.2.0 predates this PR. A CI user following the updated README with language: rust gets KeyError: No parser for 'rust'. Same for the default in actions/analyze/action.yml and the reusable workflow.

Proposal: merge without the version-string bumps, cut 0.2.1 (or 0.3.0 given the recovery change, if it stays) right after, then bump pins in the release PR. @lemduc — OK with that, and which version number?

Minor (non-blocking, noting for the record)

• _crate_context re-parses each crate's Cargo.toml per file — worth memoizing by directory for large workspaces.
• cache_key with language=None now does a second full rglob for Cargo.toml on every project regardless of language; collecting manifests during the first walk would avoid the double traversal.
• Hashing tests:excluded into every key invalidates all existing cached graphs on upgrade — one-time cost, deserves a release-note line.
• use foo::{self as alias} produces import path ("foo", "self"), so references through the alias don't resolve. Rare; fine as a documented boundary.
• .github/skills/harden-tree-sitter-parsers/SKILL.md and docs/BUG_CATALOG.md bring my review-process conventions into the repo (README links them). @lemduc if you'd rather keep process files out of the tree, I'll drop them from this PR.

---

🤖 Review generated with Claude Code

[lemduc]

Thanks for the careful self-review — taking your three points in order.

1. cfg(test) comment bypass — good catch, and I see the fix (plus the non-mod #[cfg(test)] leak and the file-cap removal) landed as #29 against this branch. To be explicit about sequencing: #29 hardens this parser, so it lands with or immediately after this PR — never on its own. Overall order is #33 first, then this one after a rebase onto main, with #29 riding along.

2. recover.py facade refinement — you're right that it's a cross-language behavior change hiding in a Rust PR, and your (a)/(b) framing is the right one. I want to look at the baseline-shift impact before choosing, so I'll come back with that call in a follow-up here rather than guess now.

3. Release sequencing — agreed on the shape: merge without the version-string bumps, and only move the actions/analyze / arcade-agent-version pins in the release PR itself, so the README never advertises a version that can't parse Rust. The exact number depends on where 2 lands (parser-only reads as a patch; if the recovery change stays it's at least a minor), so I'll settle both together.

On the process files (SKILL.md / BUG_CATALOG.md): let me look at those with the point-2 decision and follow up — no need to drop them preemptively.

The minor notes all look reasonable. The cache-key invalidation deserves the release-note line you suggest; _crate_context memoization and the double rglob are fine as small follow-ups rather than growing this PR further.