sx/docs/fork-c/S0.1-byte-baseline-and-commit-discipline.md

S0.1 — Byte-baseline (baseline-green only) + commit-classification discipline

Authority: runs/stdlib/design/fork-c-plan/planspec-r3.json (S0.1) + runs/stdlib/design/fork-c-deepdive/reconciled.md (§6 green-lock). Base: wt-stdlib-base @ 1f755284d98c6e8ebba953045c06e35d8cbe6278 (A–E6a merged). This is documentation only — no production code change, no behavior change.

1. The byte-identity reference

The single-author byte-identity reference that every later Fork C commit is checked against is the committed examples/expected/* snapshot set on the baseline commit above. We do not copy every file into this doc; the snapshots ARE the reference, and the reproduction is a documented zero-diff command (§2). Single-author byte identity is held structurally by nominal_id == 0 ≡ structural intern (src/ir/types.zig internNominal), which S1–S2 keep additive and S3 preserves through ordinal-0 materialization.

The baseline-green corpus that the reference covers:

segment	what	how it is exercised	count @ S0
baseline-green examples	every examples/<name>.sx with an active examples/expected/<name>.exit marker (incl. the 6 harvested baseline-green cases 0795–0798, 0823, 0828)	bash tests/run_examples.sh	540 active markers
FFI corpus	examples/12xx–14xx (96 entry trees; 95 with active markers; ~418 files incl. module/.c/.h/.m companions)	same runner (markers)	95 active markers
LSP completion/hover smoke	LSP unit tests under zig build test — analyzeDocument flat/namespaced import + the lsp corpus sweep: every examples/*.sx analyzes without panicking sweep + definition/references/inlayHint	zig build test	—

Count note: reconciled.md §1 cites "116 files in examples/12xx–14xx". The live tree has 96 entry .sx trees (95 with active markers); the "116" is a stale historical figure. What is load-bearing is the invariant — all FFI 12xx–14xx examples stay byte-stable — which run_examples.sh enforces via their markers, not the exact historical count.

2. The zero-diff reproduction method

tests/run_examples.sh runs sx run <entry> for every <name>.sx that has an expected/<name>.exit marker, normalizes stdout/stderr identically for expected and actual (address hashing → 0xADDR; absolute …/examples|issues/ paths → repo-relative), and diffs exit + stdout + stderr (+ optional .ir). A zero-diff run is the byte-identity check:

export PATH="$HOME/.zvm/bin:$PATH"
zig build                       # build the compiler under test
zig build test                  # unit tests incl. the LSP completion/hover smoke + corpus sweep
bash tests/run_examples.sh      # must print "<N> passed, 0 failed, 0 skipped*, 0 timed out"

Any later commit whose single-author output drifts shows up as a FAIL with a printed diff. tests/run_examples.sh --update regenerates the snapshots — it must only be run intentionally when a commit's classification (§3) authorizes an output change, and the diff reviewed. (* skipped counts .exit markers whose .sx is absent; it is 0 on a clean tree.)

3. Resolver-target EXCLUSION (recorded, not silently absent)

The resolver-target set is excluded from both the byte-baseline AND the active run_examples.sh set, because it encodes known-wrong old behavior (silent resolve / exit 0 / under-diagnosis / wrong author) — a stale old-selector verdict must never enter the baseline as if it were correct. The exclusion is explicit and listed, not absent:

• the 17 harvested 08xx resolver-target cases (0811, 0812, 0813, 0814, 0815, 0816, 0817, 0818, 0819, 0820, 0821, 0822, 0824, 0825, 0826, 0827, 0829) + the re-filed E6BR-5 regression;
• enumerated in tests/resolver-target/manifest.md, with TARGET goldens in tests/resolver-target/expected/, held inactive (no examples/expected/ marker) and asserted currently-failing by tests/resolver-target/run_resolver_target.sh;
• full disposition in S0.2-e6b-disposition-and-two-corpus-partition.md.

They flip to active + green at S3.9 and only then join the baseline.

4. Commit-classification discipline

Every future Fork C migration commit (S1→S6) is tagged with exactly one of three classes, stated in the commit subject/body so a reviewer knows what byte-effect to expect:

tag	meaning	expected byte-effect on the baseline-green corpus
mirror	builds new facts / a new resolver path in parallel, while lowering still consumes the old path (S1–S2; the assert-only Debug mirror)	zero — single-author output byte-identical; provably zero byte-risk
consumer-cutover	switches a consumer from the old path to the resolved facts (S3 materializer / calls / consts / protocol-registration / #using; S5 LSP)	zero on baseline-green — byte-identical by ordinal-0 materialization + payload-preserving facts; the only commits that may change resolver-target (the S3.9 flip is a cutover)
deletion	removes a now-dead artifact (old name selectors, *_by_source mirrors, type_bridge, findByName, the grep gate, the S2 mirror)	zero — the deleted code had no live readers after its cutover; a surviving reader fails to compile

Rules:

• A commit is exactly one class; a cutover that also deletes its now-dead source is still a consumer-cutover if the delete is the same atomic cutover (e.g. S3.10 removes the last old selector and the S2 mirror in the cutover commit).
• mirror and deletion commits MUST be byte-zero on baseline-green; if a mirror commit changes a byte, it was not actually parallel — stop.
• Only consumer-cutover may legitimately change output, and only the resolver-target corpus (never baseline-green) — that is the S3.9 flip.

5. Acceptance (S0.1) — self-check

• ✅ Byte-baseline of all baseline-green examples + FFI 12xx–14xx + LSP smoke captured and reproducible via the documented zero-diff command (§1–§2); the reference is the committed examples/expected/* at the baseline commit, re-checked by a zero-FAIL run_examples.sh.
• ✅ Resolver-target set explicitly excluded from the byte-baseline AND the active run_examples.sh set, and recorded/listed (§3) — not silently absent.
• ✅ The mirror | consumer-cutover | deletion classification rule is written (§4).
• ✅ zig build && zig build test && bash tests/run_examples.sh green over the baseline-green corpus; no behavior change (S0 adds no production code).