[kvlonge]

Why

The trusted 0.156.0-vs-0.155.0 release comparison fails only on peak memory, for three commands: lineage hub downstream and lineage rollup upstream (80 → 131 MiB, 1.63×) and dense compile (warm cache) (139 → 185 MiB, 1.33×). The same commands are 2.7× and 2.6× faster. The same lineage growth also puts the 3,000-model fresh-process lineage guards at 130–136 MiB against their absolute 128 MiB cap, so they fail on runner noise. The owner approved a one-off exception sized to the failing commands, to be removed after this release so the next one is held to the normal limits against 0.156.0.

Changes

• Peak-memory ratio limits: 1.75× for the two lineage commands and 1.45× for the dense warm compile.
• Removes the earlier temporary exceptions for lineage column trace, dag --json and the one-model edit compile, which now pass at 1.05×, 1.06× and 1.12×.
• Raises the four 3,000-model model-lineage fresh-process RSS caps from 128 to 160 MiB, marked temporary.
• Updates the release-performance tests that pin the configured limits.

Verification

• pytest tests/unit/scripts/release_performance tests/e2e/scripts/release_performance -n auto --dist loadfile: 62 passed.
• Ruff check and format clean on the touched files.