every row, every leg: the gate that keeps the edge cases held

[gafferongames]

The silence

The rows of docs/FIXED-FORM-VERSIONING-TESTS.md are the fixed form's versioning edge cases — one definition change per row, each read two ways — and each leg's harness probes them by row name: {row: "field_append"} in the Go-driven harnesses, field_append_case() in the C++ reference. CI's full lane runs all of them.

Nothing said a row is probed on every leg. A row could be written into the page, probed on four legs, and nobody would see the other five. Glenn, today:

"The edge cases must be probed, tested and known to hold, and stay held."
"Everything you decide, every edge case you fix, this creates new tests that enforce that decision moving forward across all languages."

The gate

compiler/everyrow_test.go, two tests and a ledger.

TestEveryRowEveryLeg parses the page's row tables — a table whose first header cell is row, which is the main table of definition changes and the divergence table, and not the tables about the rows — greps each of the nine legs' harnesses for each row name, and fails naming every (row, leg) pair nobody probes. The match is anchored on both sides and allows only the reference's _case / _hostile_case suffixes, so string_grow takes no credit from wstring_grow and fixed_I_grow none from fixed_I_grow_element. A leg whose harness file does not exist yet probes nothing, which is the true state of the Java leg.

A row whose NEW-READS-OLD column is — is LOCK-only: no file exists for a leg to read, its proof is the lock's refusal in internal/lockfile, and the nine read harnesses do not owe it. That is read off the page rather than listed in Go, so a row that gains a read column starts being owed on nine legs without anyone editing this file.

docs/FIXED-FORM-OWED-ROWS.md is the exception list, and it is a ledger rather than an escape — row | leg | owed by | since. The gate also fails when a line is for a pair that is probed, for a row the page does not have, for a leg outside the nine, or for a row the page marks LOCK-only. The list can only shrink. A row added to the page is owed on nine legs the minute it is written; a row that lands on a leg is one line deleted.

TestEveryRowHasCorpusBytes is the second gate: a row with a read column needs bytes, and the dump writes one row= per row into build/fixedform-corpus/manifest.txt. The corpus is a build product, so the test skips when it is absent and fails under SCHEMA_REQUIRE_CORPUS. Its exceptions are the same ledger, under the target corpus.

The counts, seeded with today's true state

rows in the page's row tables	39 (35 definition changes + 4 divergence rows)
LOCK-only (NEW-READS-OLD is —)	10 — field_undeprecate field_modify array_shape text_kind range_added default_change keyword_change closure_plain_table closure_variable_kind closure_self
rows owed a read on each leg	29
pairs (29 × 9 legs)	261
probed today	192
owed today	69 leg pairs, plus 5 rows the dump writes no bytes for
of those, unassigned	50 (45 leg pairs + 5 corpus rows)

The 69 are three groups:

• 24 — the Java leg, owed by java leg: §5, the fixed table reads backward (the lineage, LOAD's eleven steps, the two columns of every row) #920. internal/codegen/javatable/fixedversioning_test.go does not exist; java leg: §5, the fixed table reads backward (the lineage, LOAD's eleven steps, the two columns of every row) #920 writes it and names exactly these rows. Those lines go when it merges.
• 9 + corpus — fixed_I_grow_element. [4]fixed(12,4) → [4]fixed(28,4), the element's I per slot. Probed on no leg, no bytes in the corpus. Its scalar sibling fixed_I_grow is on all nine, so this is a port of a case that exists. Unassigned.
• 36 + 4 corpus — the four divergence rows, writer_bound_count, refuse_writes_nothing, unknown_census, forged_ordinal_both_plans (§5.8's 4, 9, 11, 12), probed on no leg and written by no dump. These are the rows whose counters are asserted EXACTLY rather than >= 1, so a leg that counts twice is caught nowhere else. The page says the reference and the dump are Johnny's and every other leg builds its fixture from the page's text and the row's manifest line. Unassigned.

Red first

By hand, all five refusals, each observed red and then green again: a deleted ledger line (field_append on java named), a ledger line for a pair that is probed, a row that does not exist, a leg outside the nine, a LOCK-only row listed as owed; and the corpus half skipping without SCHEMA_REQUIRE_CORPUS and failing with it. TestEveryRowProbeNameIsAnchored pins the near-misses as a unit test.

Where it runs

ci.yml's go-test, as a named step after make tables-go-versioning — that step is what builds the corpus, so the corpus half runs for real instead of skipping. It rides the plain go test ./... too, where the corpus half skips by design. Seconds: two file reads and a regexp. #966's fast lane has not merged; when it does, this belongs in lint-fast.

gofmt -l . clean; go test ./internal/ci/ green; go test ./compiler/ -run 'EveryRow|Owed' green with the seeded ledger and the corpus built.

🤖 Generated with Claude Code