fix(session): port resize-storm coalescing and owner-pane gating from main (0046 M3.3)

[yicheng47]

0046 M3.3 — third task of the M3 session-hardening slice. Ports main's 38af26c (coalesce resize storms, gate size pushes to the owning pane, #373) and e78718f (emit the resize clear in-band at settle, freeze non-owner grids); port guide is main's impl 0039 at its post-e78718f state.

What ported

• Backend width storms trailing-debounce at 175ms and settle under the session-state lock, preserving prompt DB geometry; rows-only and shell resizes stay synchronous; a real width change settles to a single in-band clear sequenced before the SIGWINCH repaint bytes (consumer seq gate closes the broadcast-ordering hole — comment at the emit site updated per review).
• Native GPUI half: prepaint geometry passes a zero/unplaced gate; each grouped session has exactly one owner pane — non-owners neither refit nor push sizes (terminal_resize.rs + pane-layout owner logic, with main's helper tests translated to Rust).
• main's manager tests ported (resize + settle suites); new native geometry/owner tests.

Explicit deferrals (no native counterpart)

React-only hidden PersistentSurfaces resize, resizeDisabled, and frontend telemetry — recorded with reasons in the impl-log entry.

Also in this PR

• docs: decision 7 revised — crate renames (runner-app → runner-backend, runner-native → runner-app) move from Phase 6 cutover to right after M3.4, so M4's UI code is written under the final crate layout.

Verification

• make verify green: 439 runner-app tests, all native tests, clippy -D warnings, fmt-check, 10/10 terminal fixtures.
• Codex-peer review clean; reviewer re-ran the gates independently and traced the debounce state machine through the kill/settle/prune interleavings.

🤖 Generated with Claude Code