Integrate all fork features and current upstream changes

[EricRasputin]

What changed

Integrate every outstanding fork branch and current upstream work into main, preserving the original commits and merge history:

• Remove worktree-opening toast during session navigation #2: remove the worktree-opening toast during session navigation while preserving restoration and error handling.
• feat: show upstream issues and PRs in Inbox for forks #4: discover both fork and upstream repositories for Inbox; keep item operations and caches scoped to the actual repository.
• Generate AI worktree names with fallback and retry #5: generate concise AI worktree names, with a stable fallback and explicit Retry.
• Upstream through e7d5623: OpenCode attachment recovery, progressive Inbox mounting, and provider token metrics.
• Record the already-present upstream Windows GitLab fix branch in history without changing its implementation.

All source branches merged without textual conflicts. All seven worktrees were clean, and every local/remote branch tip was verified as an ancestor of this integration. Three feature worktrees had no commits beyond the original main.

Why

Consolidate the outstanding work into a validated main, preserving both fork features and upstream changes in shared files.

Closes #3.

Integration corrections

Review found two behavior gaps, now covered by regression tests:

• A failed Git branch rename with an intact original checkout marked naming permanently skipped. Keep the saved suggestion retryable after a verified non-mutating live failure; interrupted-operation recovery remains conservative. The regression holds a real packed-refs lock, releases it, and verifies Retry uses the original suggestion.
• Codex and Pi/OMP replaced turn usage with the final model request. Codex now subtracts the pre-turn cumulative baseline and captures usage replay on resume; Pi/OMP replace per-assistant snapshots and sum the requests. Repeated events do not double-count, and the next submitted turn resets accounting. Context meters retain their request-based semantics. When a resumed Codex session has no verifiable baseline, omit turn spend rather than display lifetime or partial totals.

Windows CI also exposed two pre-existing Clippy failures in worktree setup. Box the large setup-operation enum variant and remove a needless return; the corrected native code passes local Clippy and all 346 Rust tests.

The Windows test run additionally caught a worktree protection mismatch for nonexistent nested paths: canonicalization added a verbatim prefix only to the existing parent. Normalize both Windows paths through the existing path conversion before component comparison. The existing shared-session protection test exercises this case. Test repositories now disable automatic CRLF conversion locally, and Windows uses a space-containing external path while Unix retains newline-path coverage.

Added a manual trigger to the existing CI workflow after automatic PR events failed to start checks. Validation runs on the final commit, with branch protection unchanged.

Standards

No documented-standard violations. One nonblocking maintenance observation remains in the upstream change: FX and Grok duplicate ACP token-metrics parsing. A shared parser can be considered separately.

Spec

Both identified requirements gaps are fixed and re-reviewed. No remaining concrete mismatches found. Primary sources were the original PR descriptions, linked issue #3, naming design documents, upstream commit messages, and upstream PR hardbeat920#209.

UI

Combines the UI changes documented in #2, #4, #5, and upstream hardbeat920#209. No additional UI redesign.

Validation

• npm run check: passed; 191 web test files, 2,096 web tests, and 346 Rust tests. Includes TypeScript, rustfmt, and Clippy with warnings denied.
• cargo check: passed.
• npm run build: passed; existing bundle-size warnings remain.
• git diff --check: passed.
• Focused regression tests reproduced the native Retry and adapter accounting failures before their fixes, then passed afterward.
• Final-commit GitHub CI: https://github.com/EricRasputin/monocode-eric/actions/runs/34853766029 (passed on macOS, Windows, and Linux at commit 2fdcded3329f9b141acad4dbb19c5203d90be685).

Checklist

• Full local automated checks complete
• All outstanding work included with original commit history
• Review findings corrected and covered by regression tests
• Final-commit GitHub CI complete

CI activation fix and merge status

Merged normally into main at 9f5ed69398a87953142bf49c6d85d78111bf25ab. Local main is synced with origin/main; all seven worktrees are clean and included. Original PRs #2, #4, and #5 are marked merged, and issue #3 is closed. All three platforms passed the normal pull_request CI run on the exact head commit.

The missing automatic checks were traced to GitHub's fork activation gate. The authenticated Actions page displayed “Workflows aren't being run on this forked repository”, even though the repository permissions API reported enabled and the individual workflows reported active. This corrects the earlier suspicion of a generic status-reporting outage.

Reapplying enabled Actions permissions through the API and reopening the PR activated normal pull_request CI. Run https://github.com/EricRasputin/monocode-eric/actions/runs/34853766029 passed on macOS, Linux, and Windows, and the required macOS check is recognized by the PR. gh pr checks --required now passes. All branch-protection settings remain unchanged; no administrator override was used. The merge also automatically triggered push CI on main.