Re-baseline the upstream surface watch at Claude Code 2.1.300; hold cached-artifact checks to the baseline - #1155

…d cached-artifact checks to the baseline

The command center's re-baseline follow-up, approved 2026-10-15T01:48Z. Four
CachedArtifactTests pinned counts read from an older cache (SDK 0.3.289, Codex
rust-v0.160.0) and failed against the host's watch cache at 2.1.300.

Re-baseline: `python3 -E -s -B scripts/upstream_surface_watch.py --network
--write-baseline --force`, run 2026-10-15T06:38:40Z under a scratch
XDG_STATE_HOME. Every source came from the network (coverage.from_cache empty),
and the host's watch state directory was left unchanged (mtimes compared).

Watched versions, baseline before -> after:
- Claude Code, npm latest: 2.1.296 -> 2.1.300 (dist-tags: stable 2.1.292,
  next 2.1.300)
- @anthropic-ai/claude-agent-sdk: 0.3.296 -> 0.3.300 (claudeCodeVersion
  2.1.300, matched)
- Codex stable release rust-v0.162.1 and binary codex-cli 0.162.1: unchanged

Counts that changed:
- claude:env: 417 -> 422 (+5 names)
- claude:setting from settings-reference.md: 171 -> 172 (a key sdk.d.ts
  already typed; the union stays 184 and sdk.d.ts stays 174)
Unchanged: claude:hook 33, claude:mod 6, codex:config 1057, codex:feature 167.
No name was removed and no feature stage changed. The baseline no longer
records the claude-doc-docs-index and claude-doc-platform-release-notes fetches,
because no enabled doc row reads those pages any more.

Not dispositioned here (the dispositions catalog is a separate review). Four
of the five new env names have no row, and the new baseline now holds them, so
daily diffs will not list them again:
- claude:env:CLAUDE_CODE_ARTIFACT_PUBLISH_EDITS
- claude:env:CLAUDE_CODE_DISABLE_FILE_MENTIONS
- claude:env:CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS
- claude:env:CLAUDE_CODE_YOU_SHOULD_KNOW_DEBUG
The fifth, claude:env:CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS, already has a
row. The claude:doc:memory and claude:doc:skills bodies differ from their
reviewed digests and stay unreviewed. The dispositions catalog's
baseline_versions still names 2.1.296.

Tests: CachedArtifactTests now hold the parse of each unmodified cached
artifact to the committed baseline instead of a release-specific literal (173,
(33, 173), 171, 1034, and the docs-only 10 of two more tests in the class).
When the cached body is the artifact the baseline recorded (same sha256), the
names must equal the baseline's. For a per-source parse, the count must match
and the names must lie within the kind. When the cache holds another release,
the count must keep the watch's own 80% floor (usw.floor_failure), and at least
one docs-only key must exist. The comparisons with reformatted, escaped and
corrupted copies are unchanged. One test loses "ten" from its name, now
test_closing_markers_on_the_docs_only_headings_preserve_the_real_settings.

Fail-first, at f165498 against the host cache with no
UPSTREAM_SURFACE_TEST_CACHE: four failures (172 != 171; (33, 174) != (33,
173); 174 != 173; 1057 != 1034). After: all six CachedArtifactTests pass
against the host cache and against the scratch cache the baseline was built
from.

Mutations, in a scratch copy of the watch script, against the host cache:
- u-escape decoding removed from ts_string_value: 2 failures (both escape tests)
- ATX closing hashes no longer stripped: 5 errors (both closing-marker tests)
- one sdk.d.ts key lost on every parse: 4 failures (173 != 174; docs-only
  11 != 10), which the escaped-copy comparisons alone would miss
- unresolved $ref ignored: 4 failures; one schema path lost: 1 failure
With the base commit's 2.1.296 baseline (floor branch for sdk.d.ts and the
settings page), the unmutated tests pass, and the lost-key mutant also passes:
the floor tolerates a 20% loss.

docs/upstream-surface-watch.md no longer ties these checks to SDK 0.3.289 and
Codex rust-v0.160.0; its line count is unchanged.

Checks: python3 scripts/validate.py exits 0 after registering the three
changed files in manifests/evidence.json. python3 -m unittest
tests.test_upstream_surface_watch tests.test_currency_due
tests.test_currency_due_notice ran 233 tests, OK.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…nvariant

The command center's FULL read of #1155 at bf9153a (PASS, one fold): the four env names the 2.1.300
re-baseline added (CLAUDE_CODE_ARTIFACT_PUBLISH_EDITS, CLAUDE_CODE_DISABLE_FILE_MENTIONS,
CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, CLAUDE_CODE_YOU_SHOULD_KNOW_DEBUG) had no disposition row, so the
baseline grandfathered them out of unreviewed and currency_due (upstream_surface_watch.py:1597-1600). Each gets a
baseline-unreviewed row (the catalog's value for this case, upstream_surface_watch.py:210-220,
docs/upstream-surface-watch.md:192-193); #1131's rebase turns them into rulings.

test_every_name_a_re_baseline_added_has_a_disposition_row compares the baseline with the 2.1.296 baseline at
f165498 (from git history) and requires a row for every added name. Names that predate that baseline (1,566) are
grandfathered under the earlier rule, so the invariant covers re-baseline additions, not every baseline name.
Fail-before at bf9153a: the four names. After: tests.test_upstream_surface_watch, test_currency_due and
test_currency_due_notice OK; --check-dispositions valid (367 rows); validate rc 0.

Co-Authored-By: Claude Opus 5.5 <[email protected]>