golem15/msd-core
4
0
Fork 0
Code Issues 1 Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix(#4434): win32 unmeasured test files weigh the documented ~2.2x Windows-cost floor, not the Linux-measured mean (#4903)

Browse Source 
* fix(#4434): win32 unmeasured test files weigh the documented ~2.2x Windows-cost floor, not the Linux-measured mean

test(#4434): failing-first — win32 unmeasured-file weight must use the
documented Windows-cost multiplier, not the plain table mean

fix(#4434): makeFileWeigher takes a platform and applies
WINDOWS_UNMEASURED_COST_MULTIPLIER (2.2) to the unmeasured-file fallback
on win32 only; measured files and other platforms are unaffected

chore(#4434): changeset fragment

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* test(#4434): add fast-check property coverage for the win32 unmeasured-file weight multiplier

CLAUDE.md requires a fast-check property test for budget-limit logic; the
prior example-based #4434 tests didn't satisfy that. Adds two seeded,
bounded property tests: the unmeasured-file fallback matches the platform
rule for any measured table, and a measured file's weight is
platform-invariant for any measured table.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* chore(#4434): backfill changeset PR number (4903)

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* fix(#4434): the Windows unmeasured-file multiplier must not apply when there is no timings table at all

Real Windows CI on this PR caught a regression my own linux-only gsd-test
verification couldn't see: makeFileWeigher applied
WINDOWS_UNMEASURED_COST_MULTIPLIER even when `timings` is null (missing/
corrupt/empty table), breaking the pre-#2456 "no table degrades to uniform
weight 1" invariant several existing tests depend on. The multiplier now
only applies to a file absent from an otherwise-loaded table — the actual
#4434 mechanism (a real, loaded, Linux-measured table with unmeasured
entries) — never to the no-table-at-all path.

Six pre-existing tests that called makeFileWeigher/pack helpers with no
explicit platform (silently inheriting whatever OS runs them) now pin an
explicit 'linux' platform, since they test the platform-agnostic mean-vs-
median and Object.prototype-safety invariants, not #4434's Windows
behavior. One subprocess-level test is isolated from the real committed
tests/test-timings.json via RUN_TESTS_TIMINGS_FILE, matching this file's
existing isolation convention for cost-profile-sensitive assertions.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

---------

Co-authored-by: sim <sim@local>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
Tom Boucher
2026-09-20 20:08:53 -04:00
committed by GitHub
parent ccfed63355
commit 029acd9158
3 changed files with 143 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
.changeset/sunny-finches-wave.md Normal file
View File

@@ -0,0 +1,5 @@
---
type: Fixed
pr: 4903
---
**Windows conformance-tier CI no longer blows its own chunk timeout on unmeasured test files.** `scripts/run-tests.cjs` now weighs a test file absent from `tests/test-timings.json` at the documented ~2.2x Windows-cost floor instead of the plain (Linux-measured) table mean, on win32 only — a cluster of unmeasured conformance-tier files could otherwise pack into one chunk and exceed the 600s per-chunk backstop even though the chunk looked in-budget. (#4434)