[kvlonge]

Why

Fresh-process --no-cache compilation was doing cache-only work and retaining avoidable Python front-end overhead. The previous cold budgets were loose enough to accept the reported 5k/10k regression signal.

Changes

• Skip cache-only closure and compact-plan construction when caching is disabled.
• Replace per-reference qualification regexes with one gated scan.
• Reuse static model metadata and declaration visibility projections.
• Batch SQL-test header preparation and use faster JSON boundaries with compatibility fallback.
• Tighten 3k/5k/10k wall guards to 11s/15.5s/29.5s, with RSS and semantic-fingerprint assertions.

Verification

• Cold fresh-process guards: 3k 8.64s, 5k 12.34s, 10k 23.06s; fingerprints unchanged.
• 5k cache lifecycle: cold 15.81s, warm 8.32s; leaf, macro, and project-config invalidation passed.
• Focused regressions: 182 passed; final affected-boundary suite: 170 passed.
• uv sync --all-extras; make check-ci.
• Public-repository hygiene scan.
• Independent review completed; all four findings fixed and verified.