rf-mcp now runs on fastmcp 3.x (verified: rf-mcp 0.31.2 imports + boots its
HTTP server cleanly under fastmcp 3.2.4; RoboScope uses rf-mcp only transitively,
out-of-process over HTTP, and never imports fastmcp directly).

- backend/pyproject.toml: fastmcp <3 -> >=3.2.4,<4, rf-mcp >=0.30.0 -> >=0.31.2
  * >=3.2.4 closes CVE-2026-32871 (SSRF/path-traversal), CVE-2026-27124
    (OAuth-proxy confused deputy) + the cmd-injection, and skips the
    3.0.0-3.2.3 auth-header-leak window
  * <4 keeps an untested fastmcp 4.0 out of fresh-resolution builds
- backend/uv.lock: regenerated -> fastmcp 3.2.4, rf-mcp 0.31.2; drops fastmcp
  2.x baggage (redis, fakeredis, prometheus-client, pydocket, diskcache, ...)
- project-context.md: updated the fastmcp pin note (floor + <4 guard)
- .gitignore: ignore local-only presentation/ deck (licensed fonts + internal SVGs)

Verification: rf-mcp HTTP boot OK under 3.2.4; backend AI suite 187 passed; no
.py source changed (no lint/type regression). Reviewed by 3 adversarial agents.

Closes #35

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

The extension carried both package-lock.json and yarn.lock; yarn warns about
mixing package managers and Dependabot scanned both (duplicate alerts). The repo
is npm-based everywhere else (frontend, e2e) and no CI uses the extension yarn
lockfile.

- Remove extension/yarn.lock and the yarn-only `resolutions` block (npm
  `overrides` for lodash/serialize-javascript remain authoritative).
- Repo is now npm-only across frontend/extension/e2e (backend uses uv).

Verified: npm audit = 0 vulnerabilities; 64 mocha tests pass.

Refs #35

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

… axis, localized + auto-applied AI fixes

Flow Editor:
- Add EXCEPT / FINALLY palette items and scaffold TRY → EXCEPT → END so a
  TRY block is valid Robot Framework the moment it is added (a bare TRY/END
  was a syntax error and could not be completed via the UI).
- E2E coverage: build IF/ELSE and TRY/EXCEPT through the UI and assert they
  land cleanly in the Code tab, on disk after Save, and across a reload.

Stats:
- Success-rate-over-time chart now renders one slot per calendar day between
  the first and last data point; days without executions keep their width
  and show a faint baseline instead of collapsing the gap (even time axis).

AI failure analysis:
- Run the analysis in the frontend's current i18n language (prose translated,
  code/keywords/patches left verbatim).
- Offer one-click auto-apply of suggested unified-diff patches; the backend
  applies context-first and refuses (422) when hunks no longer match, so a
  stale patch can never corrupt a test.
- Scope the shared analysis to its execution/report so it no longer lingers
  when switching runs; discard it on leave.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>