[maxwellE]

Summary

prehash_rule hashes a rule's attributes as soon as its query batch is decoded and then keeps only tags and visibility with Vec::retain. retain drops the removed attributes' contents but keeps the vector's buffer. Attribute is 464 bytes and every queried rule carries dozens of them (all implicit attributes are in the proto), so each rule pinned tens of KB of dead capacity until hashing finished. On the large synthetic workspace that was about 420 MiB of the process's 665 MiB peak.

The fix calls shrink_to_fit() after the retain (which also covers the existing empty case). The existing prehash test now asserts capacity equals length.

Hash output, dependency edges and impacted sets are byte-identical on every workload.

Performance (make perf-proof)

Verdict: IMPROVED, no regressions. Linux, 2 vCPU / 7.8 GiB, Bazel 8.8.2, 5 interleaved iterations, --builder cargo (the --config=release build needs hosts this environment cannot reach). Base 6177715 (71.0.0), candidate tree 87ef848.

workload	stage	bazel-diff peak RSS	bazel-diff CPU
small (2k rules)	generate-hashes warm	103.3 → 93.1 MiB (-9.9%, p=0.008)	unchanged
medium (7.5k rules)	generate-hashes warm	291.3 → 175.1 MiB (-39.9%, p=0.008)	0.205 → 0.188s (-8.5%, p=0.008)
medium	generate-hashes --depEdgesFile	289.3 → 173.2 MiB (-40.1%, p=0.008)	0.205 → 0.187s (-9.2%, p=0.016)
large (20k rules)	generate-hashes cold	665.5 → 245.4 MiB (-63.1%, p=0.008)	0.554 → 0.441s (-20.5%, p=0.008)
large	generate-hashes warm	663.4 → 247.3 MiB (-62.7%, p=0.008)	0.502 → 0.433s (-13.6%, p=0.008)
large	generate-hashes --depEdgesFile	667.3 → 239.2 MiB (-64.1%, p=0.008)	0.551 → 0.432s (-21.6%, p=0.008)

get-impacted-targets is unchanged (it does not query). The CPU gain is mostly system time (page faults on the retained buffers). The full harness report is perf-proof-summary.md.

Testing

• cargo test --release --lib: 253 passed
• make perf-proof: correctness table identical on small, medium and large

Claude for Maxwell

[Copilot]

🟢 Approval recommended

No active review issues remain for the prehashing attribute-capacity cleanup.

No open issues 🎉

What changed in this PR

This PR reduces memory retained during rule prehashing by releasing unused attribute-vector capacity after keeping only tags and visibility, without changing hash computation.

Changes:

• Replace empty-vector cleanup with shrink_to_fit() to reclaim unused capacity for all prehashed rules.
• Extend the existing prehash test to assert that attribute capacity equals length.

File	Description
src/​hash.rs	Reclaims unused attribute storage after prehashing and adds a capacity assertion to the existing test.

_{Review effort: Balanced}