docs(claude): re-derive the gate table's callers against bh 0.15.0, and replace the citation form that failed (phaze-g9cus)

[SimplicityGuy]

Re-derives the "Boundary → recipe" table's caller citations against the bh actually installed, and replaces the citation form that failed rather than refreshing its digits.

What was wrong

The table cited bh 0.14.0 and seven work.py line numbers. bh 0.15.0 split that module, leaving work.py at 1777 lines, so sed -n "${n}p" returned empty for all seven (1909, 2417, 3557, 2923, 2948, 2972, 3525). They did not point at the wrong code — a wrong citation can be read and recognised as wrong. They dangled: unreadable, uncheckable, and nothing about the table's appearance changed.

What this does

Re-derived all seven by reading each new call site, not by mapping numbers forward or diffing versions. No boundary's resolved recipe changed. The BLOCK → COMMAND half was re-executed against the live ~/.beadhive/config.yaml on 0.15.0 and reproduces exactly; the BOUNDARY → PHASE half stays I, as the table's own discipline requires.

Changed the citation form. Each row now cites module.py::symbol plus the resolver call as written, the bh version sits in the table header, and a one-command recipe regenerates the whole boundary population. A bare line number is a pointer with no redundancy — sed -n '3557p' succeeds against any file long enough — which is this repo's ADR-numbering argument one level down. Not an argument from principle: all seven boundaries changed file and line, and all seven are still findable by symbol and phase literal.

Also in scope, because bumping a version citation without re-checking the claims attached to it is the defect this bead exists to fix:

• Three sharpenings — merge/merge-main is a post-land re-test with rollback, postland is molecule-only, union can only fire on a per-bead merge.
• Group-merge sits outside the negative control. It resolves the bare validate_cmd, no work.validate key can reach it, the schema accepts an inert batch:/check: key, and a batch of parent-less beads lands on main past merge-main. Escalated to HQ as hq-f9w.
• work.validate_precheck no longer exists in 0.15.0; work.always_run documented in its place.
• Ledger field list corrected to seven plus a conditional eighth, framed as two of the seven are the key and five are not, and stated as a property of the entry rather than of replay — with the command that answers it, and a note on why --git-common-dir is load-bearing.
• Stopped re-enumerating the ledger's fields in two further sections (separate commit, droppable on its own). Both listed six fields, one as an emphatic "Take from the ledger only what it stores". Pointers replace the lists; the arguments around them are untouched.
• A merge-main bounce that ran zero tests is an unmeasured gate, not a red one — cited as a dated observation, with the general form: a gate is RED only if its own pytest summary line says so.

Evidence

Three gate runs, each forced by a base move rather than by rework. The last, quoted from the log body:

phaze test database: 'phaze_bhcite_bcfa3043_test' on localhost:5433 (from TEST_DATABASE_URL, exclusive)
🎯 selector: escalate 1 changed file(s) have no coverage, no test reaching them and no paired test: CLAUDE.md
== 8038 passed, 3 skipped, 180 deselected, 96 warnings in 1264.05s (0:21:04) ===
TOTAL                                               18199    169   3922    141  98.56%
Required test coverage of 95.0% reached. Total coverage: 98.56%
GATE_EXIT=0

Submit replayed that verdict rather than re-running it:

✓ validation verdict reused (sha 4976dfc, tree e1b24a2, recorded 2026-08-25T12:33:31-07:00)

Reviewer notes

• The verbatim operator question and two dated measurements keep their 0.14.0 citations, with re-verification appended rather than substituted. Editing a quoted question so the asker appears to have said something they did not would falsify the durable record; a measurement is a property of a run at a time. Correct a standing claim, preserve a dated one.
• docs/design/0016-transferred-model-verification.md §3.3 describes this table's pre-repair state and explicitly defers the fix here ("Tracked as phaze-g9cus; do not repair them from here"). Its Consequence row reads present-tense and will describe a state that no longer exists once this lands. Left alone deliberately — an ADR is a dated analytical record, and it points at this bead — but flagging it for whoever next touches that file.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GazEx6kCcyQuRvifWb3aBF