The documented error table is held to ERROR_RULES

[JArmandoAnaya]

docs/api.md's "full table" is where a client author reads the error-code inventory, and nothing
held it to the code. It fell nineteen codes behind before anybody noticed.

The sync half was already delivered

Measured at the merge-base before writing anything, rather than taken from the issue body:

ERROR_RULES codes: 75      table codes: 80
In ERROR_RULES, NOT in table: []
In table, NOT in ERROR_RULES: INTERNAL_ERROR, METHOD_NOT_ALLOWED, NOT_FOUND,
                              UNAUTHORIZED, VALIDATION_ERROR
Status disagreement: []

The documentation audit #533 (d723c4b, 2026-08-10) synced the table and the issue was never
updated. The five remaining extras are exactly the five the document itself names in prose as
raised by the framework and the auth guard. So the drift is gone and this PR is the other half
— which was always the half that mattered, since an ungated mirror is a second spelling waiting to
drift back.

Mechanism: parse the markdown in place

The issue leaves the choice open between parsing the table and promoting it to a
generated-and-committed artifact on openapi.json's footing. Parsing, because the artifact route
needs a script, a --check mode, a CI step and a regeneration ritual, for a table whose rows carry
grouping and prose a generator would flatten — and the invariant is identical either way. It is
~55 lines in a module that already imports ERROR_RULES.

Placement is tests/server/test_errors.py, beside test_the_status_and_code_of_every_error: that
test already owns the ERROR_RULES ↔ kernel/errors.py correspondence, and the document is a
third side of the same triangle. frontend/ui-core/src/tokens.test.ts is the precedent for holding
a document this way.

What landed

Three tests, both directions and the status:

Test	Fails when
test_every_error_rule_appears_in_the_documented_table	a code ships and nobody edits the document
test_every_documented_code_still_exists	a row names a code that was renamed or deleted, or sits under the wrong status
test_the_published_message_table_matches_expose_message	the 5xx opt-out table and expose_message disagree either way

FRAMEWORK_CODES names the five that have no kernel class to map (VALIDATION_ERROR,
NOT_FOUND, METHOD_NOT_ALLOWED, UNAUTHORIZED, INTERNAL_ERROR) — the document is the only
place they are written down, so they are listed rather than derived.

Two decisions worth the diff:

• Codes are collected as (code, status) pairs, not a mapping. A mapping lets one code listed
  under two statuses silently overwrite the other; as pairs it arrives as two members and the
  stale one is caught. M3 below is that case.
• An empty parse raises. Without it, deleting or reshaping the table would make
  test_every_documented_code_still_exists pass on nothing — a gate that stops guarding exactly
  when the document is being edited carelessly, which is the only time it matters. M6 is that
  case.

Also corrects server/errors.py's own module docstring, which said four errors opt out of 5xx
opacity when it has been six since the inference work landed. The 5xx table in docs/api.md
was already right; only the docstring had drifted, and it is now gated from the document side.

Test plan

Every stage's exit code, staged against the harness ceiling:

Stage	Exit
bash scripts/check.sh python (106s tests, ruff, mypy, 4/4 import contracts)	0
bash scripts/check.sh frontend generated	0
bash scripts/check.sh docs	0
bash scripts/check.sh browser (245s e2e + 81s real-server cycle)	0

tests/server/test_errors.py: 26 passed.

Mutation-verified — six cases, each turning a named test red, reverted by its own recorded
diff with a clean-tree assertion before and after:

Case	Expected red	Actually red
M1 · a code dropped from the table	test 1	test 1
M2 · the table names a code that does not exist	test 2	test 2
M3 · a code listed under the wrong status	tests 1 + 2	tests 1 + 2
M4 · an opt-out row dropped	test 3	test 3
M5 · the opt-out table gains a row	test 3	test 3
M6 · the ## The full table heading renamed	tests 1 + 2	tests 1 + 2

M1 reddened only test 1 and M2 only test 2, which is the evidence that the two directions are
independent rather than one test doing both jobs.

Found, not fixed

• The docstring count is corrected but still a bare numeral, so it can drift again the way it just
  did. Removing the count instead was considered; the correction is what was scoped here, and the
  countless phrasing is a one-word follow-up if it is wanted.
• A fresh worktree needs pnpm install and pnpm --dir docs-site install before check.sh
  reaches the frontend and docs stages. Pre-existing, unrelated to this diff.

Closes #524.