Stop Context7 serving anti-pattern samples; collapse three doc copies into one

[franco-zalamena-iterable]

Collapse to one doc corpus; drop CI that only served CI

Two commits on fix/context7-exclude-eval-transcripts:

1. 53842b7 — stop Context7 indexing the eval sample transcripts
2. 8c9091a — collapse three doc copies into iterable-android/reference/

1. Context7 was serving anti-pattern code as authoritative

The published library /iterable/iterable-sdk-skill (High reputation, under
Iterable's name) was indexing eval/transcripts/android/*.baseline.sample.md.
Those are hand-authored deliberately wrong samples — the baseline arm the eval
scores the skill against.

A query-docs for initializeInBackground returned 3 of 5 results from those
transcripts, including:

• setAutoPushRegistration(true) alongside a manual registerForPush() (PITFALLS Fix Cursor README install steps so the skill loads on first try. #6)
• setEmail inside the init callback (PITFALLS Package as a Cursor plugin (skill + Context7 MCP) #2)
• the 3-arg initializeInBackground overload that binds config to the callback slot (PITFALLS #18)
• 3.7.0 as the version to depend on

folders: ["polished"] was not being honoured — results cited sources/,
iterable-android/snapshot/, SKILL.md and PITFALLS.md too. So the fix adds
excludeFiles patterns rather than relying on folder scoping alone.

This is the only change here that affects anyone outside the repo. Worth
re-querying the index after the crawl to confirm the exclusion took.

2. Three copies of the same 14 docs

sources/ (raw fetch) → polished/ (transform output) →
iterable-android/snapshot/ (byte-identical copy of polished/).

The agent only ever read the third. The first two existed so CI could diff them
against each other, and snapshot:verify existed to catch the copy step a
reviewer would forget. That's the "nest where CI requires things that are just
there for CI" — the checks were real, but what they protected wasn't.

Now: one copy at iterable-android/reference/, written directly by
pnpm refresh:docs (fetch + transform in one pass, nothing intermediate on
disk). 12,192 lines deleted, 365 added.

Two real bugs this surfaced

The refresh workflow opened an empty PR on every run. polished_at was
regenerated on every polish, so git diff --quiet sources/ polished/ always
reported changes even when no doc had moved. The PR would open with an empty
slug list. Verified before the fix: re-polishing untouched docs produced
"14 files changed, 14 insertions, 14 deletions" — all timestamps. Provenance now
comes from source_sha, and a no-op refresh is genuinely a no-op ("0 written,
14 unchanged").

publish-context7.yml had never once succeeded — 4 runs, 4 failures,
POSTing to a placeholder URL with a secret that was never provisioned, fired on
every push touching context7.json. Deleted; Context7 crawls on its own.

What else went

• snippets: frontmatter manifest — ~12KB of hashes and line counts across
  the corpus. Its only consumer was the validator that checked it against
  itself, and SKILL.md had to tell the agent to skip past it on every read.
• check:snippets and lint:chunking — both hardcoded exit 0, so neither
  could fail a build. check:snippets pulled Kotlin 1.9.24 + Java 17 into every
  CI run to emit warnings that were skipped locally anyway (no kotlinc on PATH).
• validate-plugins.yml — folded into validate.yml, which no longer needs
  the Java setup that justified a separate workflow.
• .context7-library-id — TODO-PHASE-3/iterable-sdk-skill; nothing read it.
• snapshot.ts, fetch.ts, polish-layer-a.ts, recompute-manifest.ts.

What CI checks now

typecheck && validate:reference && validate:plugins. Every gate can fail.
validate:reference gained a check the old one didn't have: the corpus and
pipeline/config must agree on which slugs exist, so a config rename that
orphans a routing-table slug now fails instead of silently producing a dead link.

SKILL.md

• Reads from reference/; no longer claims Context7 is unpublished and
  must-not-be-called. It doesn't send the agent to Context7 either — the docs are
  already on disk, so there's nothing to gain from a network round-trip for the
  same content.
• Dropped the hardcoded "known latest: 3.8.0" floor. Maven Central is at
  3.10.1. It sat six lines under "never trust a number baked into this file".

REVIEW.md

Rewritten around the question a reviewer can actually answer — does this new
guidance contradict PITFALLS.md? — instead of the copy-and-verify mechanics
that no longer exist. If the docs start recommending something a pitfall warns
against, that's the thing worth catching, and no gate can catch it.

Verification

• Bodies of all 14 docs byte-identical before/after (diffed with frontmatter
  stripped; the only changes are the removed frontmatter fields)
• refresh:docs idempotent on a second run
• Forced a rewrite of 2 docs: bodies reproduced exactly
• validate:reference fails correctly on bad frontmatter, an orphaned file, and
  a slug listed in config but missing from disk
• pnpm check:all green from a clean clone with --frozen-lockfile
• Both workflow YAMLs parse; step sequences check out

Accepted trade-offs

• No committed raw copy to diff the transform against. Fidelity review is now
  "does this read like Iterable's docs", not a mechanical diff. source_sha still
  pins the exact upstream blob if someone needs to fetch and compare by hand.
• source_ref can vary across the corpus — skip-on-unchanged means a doc
  keeps the ref it was last fetched at. pipeline/config's source.ref is the
  latest refresh point. Documented in the schema.

Not done / open

• eval/ left alone. 519 LOC + 4 hand-authored sample transcripts, never run
  in CI, no committed results. It's tied to the beta value case rather than the
  repo's correctness, so it's your call — I didn't want to delete the evidence
  base for "does the skill help" while cleaning up plumbing. The transcripts are
  now excluded from Context7 either way.
• Version pins still say 3.7.0 — pipeline/config sdk.tag and therefore
  sdk_min_version across all 14 docs. Maven is at 3.10.1 and the corpus
  documents 3.10.0. Bumping sdk.tag rewrites frontmatter in all 14 files, so I
  left it as its own change.
• The fork franco-zalamena-iterable/iterable-sdk-skills is 20+ commits
  behind and missing .claude-plugin/plugin.json; archive or reset it.

🤖 Generated with Claude Code