fix(session): size backend-initiated spawns, paste file paths

[yicheng47]

Ships two independent fixes: #367 and #368.

Both are verified by Jason in the app; the review loop ran on the working-tree diff before this PR existed and came back clean.

#367 — backend-initiated spawns fork at 80x24 and lose scrollback

MCP mission_start and session_start_direct forked at DEFAULT_PTY_SIZE, seeding last_pty_cols = 80. The first real-cols resize then tripped the purge gate (session/manager/output.rs) and discarded the whole transcript — so a mission left unopened lost everything the agent produced before its tab was first visited. The longer it ran unobserved, the more was lost.

Fixed by caching a real measured grid in _app_state and reading it at the fork. The gate is exact equality on cols, so an estimate one column off purges exactly as hard as no estimate at all — nothing is computed from window geometry, and Rust gains no chrome constants (which would have been a third copy of the runtimeClearsOnResize duplication).

• Mission and chat panes sit in different boxes, so they cache under separate keys and never read each other.
• Empty cache still means 80x24. No fallback estimate — guessing would not avoid the purge, and the cache self-heals once any pane has rendered.
• An explicit caller-supplied size still wins; the UI path is untouched.
• The purge gate, runtime_clears_on_resize, and snapshot replay are unchanged. This removes the cause, not the safeguard.

Every resize path that reaches the backend records the settled grid, not just the main funnel — mission slot panes mount under display:none and push their first real grid through refreshActiveTerminal, which would otherwise leave the cache empty for exactly the panes this fixes. The resize dance records the settled rows, never the nudge.

#368 — pasting a copied file inserts nothing

Copying a file puts no text on the clipboard, only public.file-url, so xterm's default paste inserted nothing. Now the flavor is read through a direct objc2 NSPasteboard binding and the POSIX paths are injected instead, matching Terminal.app / iTerm2 / Ghostty.

• Gates on absence of usable text, not item.kind === "file" — whether WKWebView exposes an arbitrary non-image file as a file item is unverified, and a kind-based trigger that never fires would make the feature silently do nothing.
• Commits synchronously before consulting the pasteboard. preventDefault after an await is a no-op, so the handler cannot call the command and then decide.
• Ordinary text pastes return early and pay no IPC.
• A file reference beats image bytes: a Finder-copied shot.png pastes its path, while a screenshot or browser copy keeps bug: image paste into runner terminal not captured by agent CLI #79's attach flow exactly.
• Quoted only when the path contains whitespace or shell metacharacters, with embedded single quotes escaped. Injected through raw stdin, never inject_paste — nothing is submitted.

The orchestration lives in src/lib/terminalPaste.ts with injected effects so its ordering is testable without mounting xterm.

Testing

Local, all green: cargo fmt --check, cargo clippy --workspace --all-targets, cargo test --workspace (552), pnpm exec tsc --noEmit, pnpm run lint, pnpm test (268).

New coverage: _app_state round-trip per surface incl. absent-key and pool-reopen persistence; 24 paste tests covering the text gate, quoting, and the full orchestration. The orchestration tests were mutation-checked — preventDefault moved after the await, text paste doing IPC, the #79 image-bytes fallback deleted, a trailing Enter on injection, and image bytes beating a file reference each fail the suite as they should.

Manual smoke-test passed by Jason across both docs' Verification lists.

Notes

• The two fixes share commands/session.rs, lib.rs, api.ts, and RunnerTerminal.tsx, so they land in one implementation commit — splitting them would leave an intermediate commit that does not build.
• objc2-app-kit gains NSPasteboard + NSPasteboardItem and objc2-foundation gains NSURL; objc2 gates every class behind its own feature, so these would not compile otherwise. No new packages, so Cargo.lock is unchanged.

Closes #367
Closes #368