Fast pin: tiered test suite + stabilize-fast (target: pin in ~20s)

• Type: chore (build/test infrastructure) — Track A
• Status: done
• Opened: 2026-07-03

Problem

Pinning a version today takes several minutes. Breakdown (2026-07-03, v162):

• stabilize: test — the FULL make test suite is a hard prerequisite (minutes: hundreds of compile+run+diff steps, qemu legs, dwarf smoke...).
• The fixedpoint chain itself is ~4 self-compiles x 10.4s ≈ 42s.
• pin is instant.

Regressions have been rare lately; the full gate is still valuable but does not need to run on EVERY iteration pin.

Plan

1. test-smoke: one curated aggregated regression program (or a small fixed set, <20s total) that packs the historically-regression-prone surfaces into single binaries: managed strings (COW/concat/refcount churn), dynarrays (SetLength/Copy/Insert/Delete incl. managed elements), records (managed fields, by-value args), classes (ctor/virtual/metaclass), sets, Int64 arithmetic, exceptions, write formatting, and a mini self-referential parser snippet. Assert-count style (total ok N / N) like test_dynarray_insert_delete. New features append cases here AND to their full-suite test.
2. stabilize-fast: same fixedpoint + byte-identity chain, but prerequisite = test-smoke instead of test. This is the everyday iteration pin. Full stabilize (full test) stays for milestone pins / before pushing a batch — policy note in parallel-tracks.md.
3. Wall-time target: 20s = fixedpoint chain (~42s today) + smoke. Needs the compile-speed work to land too: [[perf-compiler-hotspots-algorithmic]] (self-compile 10.4s -> ~7s) and [[feature-optimization-levels]] (2.04x codegen gap -> fixedpoint compiles ~2x faster again). With both, 4 x ~3.5s + smoke ≈ 20s. Also consider trimming the fixedpoint chain: gen2==gen3 byte-identity already proves the fixedpoint; s4/s5 re-derivations may be redundant (audit what each cmp actually guards before cutting).
4. Optional: make test parallelism (many recipe lines are independent compile+run pairs; -j unsafe today because of shared /tmp names — namespacing outputs would unlock it).

Non-goals

Weakening the full gate: make test remains the bar for stabilize-before- push-batch, releases, and anything touching codegen/ABI/ELF.

Acceptance

make stabilize-fast green end-to-end in <=60s on the dev box today (<=20s once the perf siblings land); policy documented; full stabilize unchanged.

Resolution (2026-07-03, same day)

Landed test-smoke (11 curated tests: dynarray torture + insert/delete, frozen-string reentrancy, AnsiString, class-of/metaclass ctor, exceptions, record byval, const-record prebody, mutex + TThread sync under --threadsafe) + the full self-host chain (self -> next -> fixedpoint, cmp = the byte-identity proof) producing the artifacts the recording step pins. stabilize-fast records the proven fixedpoint binary directly — the full target's s4/s5 re-derivations only re-prove what cmp(next,fixedpoint) already established, so the fast path skips exactly them and nothing else. Full stabilize (full test + s4/s5) unchanged, now sharing the stabilize-record step.

Measured: 18.0s wall — the 20s goal met same-day, because [[perf-compiler-hotspots-algorithmic]] item 1 (symtab hashes) landed first and cut each self-compile 10.4s -> 5.9s. (The plan's arithmetic assumed ~42s of compiles; hashing made the whole chain ~18s with tests included.) Policy documented in parallel-tracks.md. make test parallelism (item 4) not needed — left as a note, not done.