feat(e2e): live-client smoke test harness (Phase 1 + Phase 2) - #45

Phase 2 built the read/write dual-connector pattern for google-docs only.
Generalize it to every server under src/: google-drive, google-gmail,
google-calendar, google-sheets, google-slides, clickup, slack, slack-user.

- tools.ts -> tools/<service>.ts + barrel. Lists are derived from each
  server's `annotations.readOnlyHint`; the derivation reproduces the
  hand-curated google-docs split exactly, which is what validates it.
- promptTemplates: per-service connector pairs, `awesome-mcp-<service>-<mode>`.
  runSmokeTest now prepends preface(service, mode) itself so a test cannot
  name the wrong connector.
- Per-backend direct-API clients + scratch factories. googleClient gains the
  slides/gmail/calendar namespaces (same OAuth plumbing); clickup and slack
  get their own thin clients. Every sweep is double-scoped: container plus
  the shared `[e2e]` name prefix.
- 18 smoke tests (1 read + 1 write per service).
- test:gate is an explicit manifest, not a glob. GATE is keyed on
  WriteTestedService, computed from the WRITE_TOOLS arrays, so a new
  write-capable service fails typecheck until it is gated.
- Wire the new secrets/vars through both e2e workflows.

Also fixes a pre-existing bug: `--test "tests/**/*.smoke.ts"` does not glob
on Node 20 (workflows pin 20), so the nightly regression matched zero files
and reported success. test:full now discovers files explicitly.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Brings PR #45 up to date with main, which had moved ahead by five commits
touching the same files the e2e harness owns.

Conflict resolutions:

- .gitignore — took main's. #45 broadened the rule to a blanket `.claude/`,
  but main has since checked in `.claude/skills/**` on purpose ("skills are
  checked in"). Main's narrower rule still ignores settings.local.json, which
  is the file that actually embeds credentials in allowlist entries, so #45's
  security intent is preserved without untracking the skills.

- .github/workflows/e2e-regression.yml — pinned actions/checkout, setup-node
  and upload-artifact to the same commit SHAs main uses in e2e-smoke.yml.
  Not a textual conflict: e2e-regression.yml is new in #45, so it never
  received 0ff451f's pinning pass and would have landed unpinned against a
  repo where 7 of 9 workflows are pinned.

Carried in from main unchanged: the claude-desktop / chatgpt-web
waitForResponseComplete race fixes, persist-credentials: false on checkout,
and the Appium --base-path removal (the runbook reference went with it).

Verified on the merge result: root typecheck clean, root lint 0 errors,
e2e typecheck clean, test:gate and test:full both resolve 18 smoke tests and
fail only on unset credentials.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The per-service tool lists omitted both shared tools for all nine services.
The pattern that generates them matches literal `addTool({name, annotations})`
blocks in src/<service>/server.ts; the shared tools arrive via
registerMintRestBearerForCurl()/registerListRestEndpoints() helpers, so it
never saw them.

That is not cosmetic. mintRestBearerForCurl has readOnlyHint: false, and
runbook.md tells operators to uncheck every tool in a service's WRITE_TOOLS on
its -readonly connector. Absent from every list, it stayed CHECKED on every
readonly connector. The bearer it mints is scoped to the USER, not to a service
and not to reads — createServiceAuth resolves the token to a user and then
looks up that user's connection per route — so a "readonly" connector could
mint a bearer reaching WRITE endpoints on every service the account has
connected. That defeats the entire point of the two-connector split.

mintRestBearerForCurl -> WRITE_TOOLS, listRestEndpoints -> READ_TOOLS, with a
comment in each file recording why they are hand-maintained, and a callout in
runbook.md's blocking step since this is the one checkbox that matters most.

Merging main in also surfaced real drift that had accumulated since these lists
were written, so all nine were regenerated from current source:
- clickup +11 tools (tags, task events, doc images)
- slack-user +7 tools (search, channel events)
- downloadFile was missing from BOTH slack lists: the extraction pattern
  required `annotations:` to immediately follow `name:`, and both Slack servers
  put an explanatory comment between them. The pattern now tolerates that, and
  the addTool-count assertion that caught it is what proved the lists complete.

Counts now: docs 30, drive 15, gmail 14, calendar 6, sheets 12, slides 6,
clickup 45, slack 8, slack-user 15 — each plus the 2 shared.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>