[SizzleUnrlsd]

Summary

Part of #69: measurements on large projects and the hot spot they revealed.

• Measured (single process, cold, bundled plugins): healthchecks, 45k lines, 9.5 s and 107 MB; rich, 52k lines, 7 s and 114 MB; wagtail, 273k lines, 123 s and 470 MB. A warm --cache brings healthchecks to 1.1 s; --jobs 4 to 8 s.
• Profiling healthchecks showed the taint engine re-collecting every analysable function of the module for each function analysed: the enclosing-function search returned None both for "top-level" and "not found", so every named function fell through to the lambda fallback, and the collection itself was never memoised. About a quarter of the cold time.
• Acceptance tests first (test: define the module function collection as computed once), then the fix: analyzable_functions is memoised per module (identity-keyed, bounded), and _enclosing distinguishes a top-level definition from an absent one. healthchecks drops from 9.5 s to 7 s, rich from 7.2 s to 5.8 s, wagtail from 123 s to 85 s, same findings.
• healthchecks (pinned) joins the regression corpus, so a real 45k-line project is analysed on every change; the usage guide gains a Performance section with the measurements.
• The measurements also surfaced precision and coverage gaps (walrus, yield from and starred targets rejecting whole files; hex digests flooding high-entropy-string; credentials in test fixtures; Django redirect("name", arg)); they are recorded as tasks on Precision: nonlocal write-back and a broader regression corpus #71.

Part of #69.