tables: a float rides as its bit pattern, pinned; the tool's cook-check carries the map slot and the list halves (#480, #380)

[rowan-claude]

Two gaps in the C++ column's contraction pass, red first, in one PR.

Closes #480
Part of #380 (the tool's cook and uncook halves for a map stay owed there, and refuseToolMaps points at it)

1. A float rides as its bit pattern (#480)

The gap. A float32 NaN payload rides as its 32 bits in every port, and the page never said so. grep -n NaN docs/SPEC-TABLES.md found the widened rung's aside about payloads being exactly representable and nothing that made the bits a RULE; docs/PORTING.md had no row; no corpus instance carried a signalling NaN. A backend holding a float32 in a wider cell quiets one through the hardware conversion, and the byte identity nine languages hold over one corpus fails on exactly those values.

The red. The tool's own cook is one such backend. Against the new pin, before the fix:

the signalling slot holds 0x7FC00001 and the pattern is 0x7F800001: the cook converted it
the negative slot holds 0xFFE5A5A5 and the pattern is 0xFFA5A5A5: the cook converted it
the wire did not survive the cook: 127 bytes back, 127 pinned

internal/tablecook/region.go wrote math.Float32bits( float32( cell.F ) ) and uncook.go read float64( math.Float32frombits( ... ) ) — a float64 cell with a conversion at each end.

The fix.

• The sentence. docs/SPEC.md §4.3 gains a wire-model bullet and its float32/float64 rows say it; docs/SPEC-TABLES.md §3 gains the paragraph beside the kind-10/11 payload row.
• The row. docs/PORTING.md M21, with its nine cells: cpp and the tool carried, the eight ports ❌ #366.
• The pin. test/tables/F1.schema and F2.schema (the same root one generation on, every float32 respelled float64 — §4's float rung). testdata/wire/tables/floats_nan.bin carries a signalling NaN 0x7F800001, a NaN with a mantissa past the quiet bit 0x7FC0DEAD, a negative signalling one 0xFFA5A5A5, two signalling array elements, and two double patterns. test/tables/main.cpp writes it and reloads it with every other golden, so the big-endian leg covers it.
• The claim and its control. test/tables/floatnan_main.cpp asserts the writer's bytes, the identity read and the widened read. make tables-float-nan-negative-control puts the hardware conversion back in TableWidenF32 through an overlay, regenerates only F1/F2 from the sabotaged emitter, and the pin goes red on the payload: 0x7ff0000020000000 read back as 0x7ff8000020000000, the quiet bit the conversion set.
• Both engines pass the pin. The Go oracle (TestOracleCarriesTheFloatBitPattern, TestOracleWidensTheFloatBitPattern) and the tool's cook (TestACookCarriesAFloatBitPattern). WidenF32/NarrowF32 are exported from internal/tablewire so one pair serves both.

The text form is unreachable by design. #480 named wire, text and cook rows; the text writer refuses every non-finite float and the reader counts a NaN token as a kind mismatch (§16.2), so there is no text row to carry, and M21 says so.

The eight ports are not this PR's. Their legs are a row on #366, and M21's cells say so.

2. #380's residue: the tool's cook-check and the list halves

The red, quoted from the tree.

internal/tablecook/check.go:212, on the cook the C++ reference wrote for Army:

node 1 (Army at offset 0): Army.squads: Squad.roster: a map slot, and `cook-check`
carries no map-slot clause yet (docs/SPEC-TABLES.md §7.4): the C++ reference reads
the cook (--lang cpp), and the tool's clause is schema#380's next PR

compiler/tableslists.go:58, through schema uncook over the same file:

unit declares an unbounded array in a table closure (...): the tool's WIRE and TEXT
halves carry the construct and `cook-check` reads one, and its COOK half does not,
so this command would lay out a region short of the element arrays rather than
refusing. The C++ reference carries the cook (--lang cpp)

The map-slot clause. The four clauses a list's slot already took are one helper now, placedArray, because the two slots are one shape and lists and maps are one population in a node's extent. On top of it the map takes the fifth: the keys read ascending with no repeat, at the key's own signedness for an integer and bytewise against the used length for a bounded string.

The list's cook and uncook halves. internal/tablecook/extent.go is the node extent in the tool: one pre-order carve walk with its writing half switched off for the measure, so the two cannot drift. Layout sizes an extent-bearing node from the instance it walked, Write fills each slot's sixteen bytes and holds the extent written to the extent measured, Uncook reads them back.

The two writers meet. make tables-lists-tool-cook has the generated SaveCook land on the tool's own bytes exactly, 352 bytes, in both byte orders, over test/tables/L1.schema — a list-bearing unit with no map in it. The CLI round trip runs beside it.

Controls. tables-maps-cook-check-negative-control (the shared containment sabotage must turn the map's test red beside the list's, and the keys' ascending clause gets its own, because a shared control would leave the fifth clause untested) and tables-lists-tool-cook-negative-control (the carve lays an array unaligned).

The page lines retired. §7.4 clause 4 loses "Until schema#380 lands this clause in the tool"; §21's maps bullet loses "which is the one piece still owed"; §21's unbounded-arrays bullet loses "the tool's COOK and UNCOOK halves are owed beside the map's".

Still owed, and unchanged: the tool's COOK and UNCOOK halves for a MAP. A map adds the sort, the entry array's key order and the two reader events to the extent the list's halves already carry. refuseToolMaps still names a map-bearing unit at those two surfaces, and the page now says that is what remains.

Goldens

make update-goldens adds exactly one file: testdata/wire/tables/floats_nan.bin. No conformance json, report or cook-write fixture moved.

The six maps headers make update-goldens once wanted to move, and the two gates (tables-json-list-walk, tables-block-zero-cost) that were red on this branch's base, were main's: #668 re-pinned the headers and #674 derives the list-walk set from the IR. Merged with main both gates exit 0; on the bare branch head they still exit 2, which is the base, not this diff.

Everything this PR does own was run green locally: go test ./..., make check, make fmt, make tables-float-nan, make tables-float-nan-negative-control, make tables-lists, make tables-lists-tool-cook, make tables-lists-negative-controls (twelve controls, each red on its own sabotage), make tables-maps, make tables-maps-cook-check-negative-control, make tables-cook-cli, make tables-cook-write, make tables-cook-open, and the wire fuzzer over the new units (0 divergences across 143,309 mutants and 44 roots).

🤖 Generated with Claude Code