[BasilZafar11]

The recall, pattern, example, and corpus tests repeatedly compiled all 79 embedded patterns for each payload. They now share a OnceLock<Scanner> per integration-test binary through tests/common/mod.rs; relaxed-pattern controls retain a separate cached scanner. The four suites named in #188 use the cache, along with the other repeated scan helpers.

The static cache confirms Scanner is Send + Sync. A concurrent test verifies shared identity and that suppressed, clean, and attack reports remain independent. Synthetic probe scanners and the performance test's deliberate fresh-compilation samples are preserved. Production code, patterns, dependencies, and the complete EXPECTED recall table are unchanged.

Local before/after timings

Windows, Rust stable, unoptimized test profile, same machine and default test parallelism, one complete run each:

Suite	Before	After
recall_test	232.56 s	1.83 s
report_roundtrip_test	1.37 s	1.00 s
perf_regression_test	10.73 s	7.65 s
pattern_relaxed_control_test	16.26 s	5.94 s
pattern_test	487.35 s	1.92 s
pattern_example_test	523.09 s	1.50 s
corpus_test	70.08 s	3.35 s
cargo test command elapsed	1794.10 s (29.90 min)	259.43 s (4.32 min)

Both commands were cargo test --locked --no-fail-fast -- --skip findings_name_the_repository_path_not_the_staging_copy, excluding the existing Windows hook path-separator assertion. Command elapsed includes compilation (baseline 61 s, candidate 5.53 s); test-binary timings above exclude it. The baseline completed with only the Windows CRLF catalogue equality failure; the candidate passed after local LF normalization, which creates no committed catalogue diff. The unchanged prefilter-equivalence suite remains a substantial runtime cost (150.50 s before, 139.13 s after). These are measured local results, not a forecast for hosted CI.

Validation and CI ceiling

• Cached suite, concurrency regression test, exact recall, clean corpus, and relaxed controls pass.
• cargo fmt --all -- --check and cargo clippy --locked --all-targets -- -D warnings pass.
• The CI timeout has been reconsidered: retain 60 minutes until a complete hosted coverage run supplies the cold-build measurement; update the workflow comment to explain the decision. Coverage still runs the full suite and should benefit from the same caches.
• Hosted CI baseline from ci: test suite needs ~40 min against a 30-min ceiling — cache the Scanner instead of rebuilding it per test case (S-01) #188: the test step took 18.1 minutes. In run 38538258849, the candidate test step passed in 4 minutes 38 seconds (08:53:41–08:58:19 UTC). Coverage passed in 4 minutes 38 seconds with 93.41% line coverage. Formatting, Clippy, benchmarks, release build, and smoke tests passed.
• The workflow's only failure was the unchanged PERF-01 gate: best of three was 213 ms against the 200 ms budget. Production code and patterns are identical to the base. This result alone does not distinguish runner variance from a pre-existing budget miss. An unchanged retry requires a maintainer: both connected integration and author credentials returned HTTP 403 for rerun.

Closes #188.

[hermanngeorge15]

Reviewed and independently verified. This is the #188 fix and it does what it claims.

Measured on macOS (arm64), same machine, same command, cargo test --locked:

branch	wall time
main	11m 02s
this PR	3m 26s

3.2x, both green, exit 0. cargo fmt --all -- --check and cargo clippy --all-targets --locked -- -D warnings are both clean on the branch. That corroborates your 29.90 -> 4.32 min on Windows and 4m 38s on CI — three different machines, the same shape of result.

On the design, the parts I checked specifically:

• tests/common/mod.rs — OnceLock<Scanner> per integration-test binary is the right granularity. Rust builds one binary per tests/*.rs, so this is one compile per binary rather than one global, which keeps the suites independent.
• pattern_relaxed_control_test.rs correctly keeps a separate cached relaxed scanner. Sharing the shipped one would have broken GATE-05.
• perf_regression_test.rs keeps its deliberate fresh-compilation sampling. That was the one place a naive cache would have silently invalidated the measurement, and you avoided it.
• tests/pattern_policy_test.rs walks tests/*.rs non-recursively, so tests/common/mod.rs sitting in a subdirectory does not disturb the test-case policy counts.

CI is red on PERF-01 only, and that is not your fault — see my reply on #188. Nothing in this PR touches production code or patterns.

One non-blocking suggestion. The diff removes the measured timing breakdown comment in ci.yml (the 18.1-minute test leg and the per-binary figures). Those numbers are the evidence for why the ceiling is 60 minutes, and they are also the "before" half of the proof that this PR worked. Consider keeping them, relabelled as the pre-#188 baseline, next to your new note.

Approving. I want the gate fixed on main first so this merges green rather than merging over a red badge.