peltier.io/results/phase0/attempts.md

Phase 0 dry-run attempt log

Hypotheses are recorded here before any patch is applied; final evidence and verdicts go to the SQLite ledger and are mirrored here.

Environment for the whole session: remote 4-core container (Intel Xeon @ 2.10GHz, kernel 6.18.5), governor/turbo not controllable, benches pinned with taskset -c 2. A/A calibration on both target workloads passed with pinning (see results/calibration/).

Deviations from SPEC (Phase 0 manual dry run)

• No nested no-network container available (no docker daemon): target code runs in this session's own isolated ephemeral container, which serves as the sandbox. Phase 1 must restore the dedicated container.
• perf unavailable (container): profiling via valgrind/callgrind (deterministic instruction counts, Ir), not cycle sampling. Ir% is a proxy for cycles%; it under-weights cache/branch effects.
• Corpus files >10 MB are pinned by committed MANIFEST.sha256 + regeneration scripts rather than committed raw (repo bloat).

Attempt 1 — comrak, class 1 (build configuration)

• Logged: 2026-07-06, before patching.
• Hotspot: whole-program; upstream already ships lto=true + codegen-units=1, so classic class-1 flags are exhausted upstream.
• Hypothesis: RUSTFLAGS=-C target-cpu=native (baseline ISA: the bench machine's Xeon) enables wider SIMD paths in string scanning (jetscii, memchr paths visible in profile) and better scheduling; expect a small win, 2-6% median speedup. No FP semantics change (integer/string workload; not a -ffast-math-class flag), so auto-acceptable if gates pass. Deploy caveat: win only valid for fleets whose ISA baseline matches; recorded in workload statement.
• Patch: none (env-level build config); candidate built to a separate CARGO_TARGET_DIR from the same pinned commit.

Attempt 2 — comrak, class 2 (allocator swap)

• Logged: 2026-07-06, before patching, after Attempt 1's hypothesis.
• Hotspot: allocator symbols sum to ~26.7% of Ir (_int_malloc 9.99%, _int_free 6.85%, malloc 3.92%, malloc_consolidate 3.48%, free 2.49%) — the class-2 precondition (≥5%) is exceeded 5x.
• Hypothesis: glibc malloc is the allocator; comrak's AST/inline parsing churns small allocations. Swapping the CLI's global allocator to mimalloc cuts allocator instructions substantially; expect 8-20% median speedup on the progit workload.
• Patch plan: 2-line change in comrak CLI main + Cargo.toml dep, inside targets/comrak/workspace only.

Attempt 3 — tokei, class 1 (build configuration)

• Logged: 2026-07-06, before patching.
• Hotspot: whole-program; upstream ships lto="thin", panic="abort", default codegen-units (16).
• Hypothesis: lto=true (fat) + codegen-units=1 lets the analysis loop (perform_multi_line_analysis, 31.7% Ir) inline across cgu and crate boundaries (memchr/aho-corasick calls visible in profile); expect 2-8% median speedup. No semantics change; auto-acceptable if golden replay passes.
• Patch: [profile.release] lto = true, codegen-units = 1 in targets/tokei/workspace/Cargo.toml only.

Attempt 4 — tokei, class 3 (allocation churn / buffer zeroing)

• Logged: 2026-07-06, before patching.
• Hotspot: __memset_avx2 8.52% Ir + malloc-family + memcpy 3.97%. Cause identified by reading LanguageType::parse: every file is read via encoding_rs_io::DecodeReaderBytes + generic read_to_end. That reader lacks a read_buf specialization, so std zero-fills every grown buffer chunk before the read lands (~221 MB memset per run), and the Vec grows without the file-size hint (realloc + memcpy).
• Hypothesis: read raw bytes with fs::read (pre-sized from metadata, read_buf-specialized, no zeroing); only files that start with a UTF-8/UTF-16 BOM fall back to the original DecodeReaderBytes path over the same bytes (identical semantics — encoding_rs_io BOM-sniffs the stream head; BOM-less input is raw passthrough). Expect 5-10% median speedup.
• Equivalence risk: BOM files must transcode exactly as before — handled by falling back to the identical machinery; differential fuzz (old vs new binary on adversarial inputs incl. BOMs, invalid UTF-8, empty files) gates the change.

Attempt 5 — tokei, class 1 stacked on accepted class 3

• Logged: 2026-07-06, before patching.
• Rationale: fat LTO+cgu=1 alone measured a real +1.6% [1.0026, 1.0262] (phase0-tokei-001, rejected: under the 2% bar standalone). Stacked on the accepted read-path patch it is part of the tree that would actually ship; the combined candidate is benched vs the pristine baseline and re-gated as a whole.
• Hypothesis: combined read-path + fat LTO clears ≥10% median vs pristine (9.9% + ~1.6%, roughly additive since they touch different costs: syscall/alloc path vs codegen).
• Patch: results/phase0/tokei-readpath.patch + [profile.release] lto=true, codegen-units=1.

Outcomes (ledger is authoritative: results/ledger.sqlite)

run_id	class	speedup median	95% CI	verdict
phase0-comrak-001	1 build-config	0.9965	[0.9826, 1.0110]	rejected-bench (null)
phase0-comrak-002	2 allocator	1.0459	[1.0333, 1.0577]	needs-human-review
phase0-tokei-001	1 build-config	1.0161	[1.0026, 1.0262]	rejected-bench (< 2% threshold)
phase0-tokei-002	3 alloc-churn (read path)	1.0990	[1.0797, 1.1271]	accepted (first auto-accept)
phase0-tokei-003	1+3 combined ship tree	1.1036	[1.0852, 1.1201]	accepted — Phase 0 exit bar met

• comrak-002 gates: 848 upstream tests green on patched tree; golden replay byte-identical; pristine ASan+LSan clean; patched ASan-proper clean; LSan flags 484 B in onig (analysis: LSan × custom-allocator reachability false positive — see case-study-comrak.md). SPEC §8 routes any sanitizer flag to human review, so that is the verdict.
• tokei-001 is a real but small effect — CI excludes 1.0 but its lower bound (+0.26%) is under the 2% bar. Correctly rejected; logged so nobody re-grinds it.
• Phase 0 exit criteria met: phase0-tokei-003 is a verified +10.4% median win (95% CI [+8.5%, +12.0%]) with full gates, plus two written case studies. Stated precisely: the median clears 10%; the CI lower bound is 8.5%.
• Still open for later sessions: comrak mimalloc (+4.6%) awaiting human review; comrak class 3 (arena the inline parser, RawVec grow 2.3% + memcpy 3.1% + from_utf8 5.1%); tokei class 5 — byte-skip table for perform_multi_line_analysis (31.7% self Ir + memcmp 8.1%), sketched but deliberately not attempted: the state-machine surface needs more verification budget than this session had left.