perf(graph): do not rebuild TestFileFilter when the session already has one

[jonathanong]

Summary

An AnalysisSession owns one TestFileFilter per root. Graph config reuses it instead of calling TestFileFilter::new → project_filters again. Standalone DepGraph::build_with_plan_and_config still uses the unprepared fallback (session=None).

Call-count tests (not value equality) lock: seeded session does not increment test_filter.builds / project_filters; unseeded session builds once; two roots do not share a cell; session=None still compiles project filters without a session build.

New bench group graph_gates_build_session (does not change the existing graph_gates_build identity).

Local Criterion (source of truth)

Machine: macOS aarch64, rustc 1.96.0, --sample-size 20. Same machine before/after. CodSpeed is not used as a gate.

Unprepared graph_gates_build still compiles test globs once per build (path unchanged; movement is noise):

bench	before (mean)	after (mean)
graph_gates_build/1	57.5 ms	52.8 ms
graph_gates_build/4	56.9 ms	46.5 ms

Prepared session path vs unprepared before on the same machine:

bench	unprepared before	session after	vs unprepared
graph_gates_build_session/1	57.5 ms	26.0 ms	−55%
graph_gates_build_session/4	56.9 ms	35.8 ms	−37% (high variance on [4])

Shepherd Journal

• Split graph_gates.rs session bench and builder.rs session entry into sibling modules to satisfy rust-max-lines-per-file (4dcd916). CodSpeed extract regression used unexpected base main and different runtime environments; local Criterion is the gate.
• Codex quota IC_kwDOSaSRXM8AAAABStTVkA/IC_kwDOSaSRXM8AAAABStTW5A and Sourcery budget review PRR_kwDOSaSRXM8AAAABMTLc_A are bot-limit notices; minimize only.
• codecov javascript "No report found to compare against" and rust "No coverage information found on head" while jobs are still inProgress are stacked-PR upload lag, not a TestFileFilter defect.
• Rejected Codex P2 PRRT_kwDOSaSRXM6fhdbC: AnalysisSession is request-scoped with one effective config (see the method comment and prepared-analysis ownership). A different config_path is a new request/session. Keying the OnceLock by root matches visible-path cells; hashing config identity on every graph-gates lookup would undo the reuse this PR measures. codecov patch/project lag is still upload/inProgress, not a missing TestFileFilter unit.
• codecov/patch/rust 55.47% is unchanged after tests finished (35 passing). Diff includes benches/graph_gates_session.rs which is not in the coverage binary; production TestFileFilter reuse is covered by files_config_session_* tests. Project javascript/rust "no report to compare" is stacked-PR upload lag against perf: probe visible/inventory paths with OsStr order, not Path components #835. No additional production code change.
• codecov/patch/rust 55.47% was missing coverage of builder_session.rs. Added files_config_session_builder_reuses_prepared_test_filter so the session DepGraph builder path is executed twice and reuses the prepared TestFileFilter.
• Codex quota comment IC_kwDOSaSRXM8AAAABStc8SQ is a bot-limit notice; minimize only.
• Benchmark (CPU graph gates) on run 33955328154 failed installing CodSpeed CLI 5.2.1 (curl: (35) Recv failure: Connection reset by peer) — transient infra, not graph-gates session code. codecov/patch/rust rose from 55.47% to 82.48% after the session builder test; remaining gap is likely uninstrumented benches/graph_gates_session.rs. Attempted rerun of 33955328154.
• Reran 33955328154 once: CodSpeed CLI install curl reset is transient infra. codecov/patch/rust 82.48% remaining gap is benches/graph_gates_session.rs (not in the coverage binary), not missing production TestFileFilter tests.