matrix: witness claude_code's prompt_submit and stop at 2.1.263

[jothimani-rajendran]

Follow-on to #113. With the stop gate's observable fixed, both remaining claimed events
were re-run against the real CLI and merged into claude_code's row by hand, the way W53
did pre_tool.

The runs

tools/experiment.py --driver "claude -p {prompt} --permission-mode acceptEdits" --agent-version 2.1.263, twice per gate (table run, then --report run), identical
results both times. Every trial reads agrees or unasserted; baseline_ok reads
unrecorded by design, being a run-health check rather than a vendor claim.

trial	field	prompt_submit	stop	asserted	status
deny	block	true	true	true	agrees
crash	fail_mode	open	open	open	agrees
transform	rewrite	false	false	false	agrees
silence	silence_means	allow	allow	—	unasserted
timeout	timeout_fail_mode	open	open	—	unasserted
unknown	unknown_verb_means	allow	allow	—	unasserted

The stop gate's deny now reads block: true ("the agent was made to continue") — the
disagreement #113 diagnosed as an instrument fault, not an agent one.

Two readings are worth naming rather than burying. At stop, transform measured "the
original input ran; the rewrite was ignored". At prompt_submit it measured "nothing ran:
the rewrite was refused or degraded to a block" — the 0/0 shape #113 stopped scoring as
null. Both mean rewrite: false, which is what the cells already asserted, but they
reach it by different routes.

The merge

• block, rewrite, fail_mode on both cells: per-claim records at live-run / 2.1.263
  / 2026-09-07, each naming the driver and gate on one line.
• Both cells gain silence_means: allow, timeout_fail_mode: open,
  unknown_verb_means: allow, each with its own record.
• Row verified.version → 2.1.263, verified.date → 2026-09-07, now that all three
  claimed events are witnessed at one build. basis stays live-run; observed already
  listed all three.
• Row narrative extended by exactly two sentences: the headless Stop block is honoured, and
  Claude Code caps the continuation loop at eight re-fires (nine Stop hook invocations in
  all, reproduced across two runs).
• to_evidence() was not used. The edit is a targeted key update, dumped with the file's
  own indent=2 + ensure_ascii settings — verified byte-stable on an unmodified round
  trip before editing, so the diff is only what changed.

Derived artifacts

data/vendors/claude_code.json (via tools/recount_vendor_config.py --write) and
examples/generated/claude_code.md (via examples/generate.py) are regenerated from the
row, not hand-edited. test_entries_match_recount and test_committed_pages_match_what_the _library_produces both require it.

One test moved with the data

test_a_recognized_field_the_cell_does_not_carry_is_unasserted read its absent field off
claude_code's stop cell, which now carries all three optional claims. It reads
post_tool instead — a cell the experiment kit cannot gate, and so one that never grows an
optional claim from a run. Still a real trial result, re-pointed at a different cell.

watch_versions: three fresh, nine unchanged

claude_code/pre_tool       fresh    2.1.263  2.1.263  0d
claude_code/prompt_submit  fresh    2.1.263  2.1.263  0d
claude_code/stop           fresh    2.1.263  2.1.263  0d
claude_code/post_tool      drifted  2.1.247  2.1.263  10d
... (8 more observation-only events, all drifted at 2.1.247)

The three claimed events read fresh, which is what this PR set out to do. The nine
observation-only events still read drifted and should: they were not re-run, their
per-claim records honestly say 2.1.247 / 2026-08-28, and making them read fresh would mean
either stamping 2.1.263 on cells nobody measured or deleting the records that pin them.
Either would turn a row-level version bump into evidence for a claim with none — the exact
failure this row's per-claim evidence exists to prevent. Left as they are.

Tests

Full suite green. Seven failures on this machine (kimi adapter, examples, git-hook exec
bit, install paths) reproduce on a clean main, are Windows-specific, and are the subject
of the companion PR. ruff check ., ruff format --check . and
tests/check_stdlib_only.py clean.

🤖 Generated with Claude Code