fix: eliminate Athena resource spikes and freeze regressions

[luckeyfaraday]

Summary

Athena's freeze/heat problem is a combination of software-composited xterm painting, terminal delivery contracts that acknowledge before xterm drains, periodic whole-corpus session parsing, and agent fleets that can bypass memory admission.

The live Surface Laptop 3 profile that motivated this PR measured:

• Electron renderer: 128.6% mean CPU (137% peak)
• Chromium Viz compositor: 61.3% mean CPU (71% peak)
• Electron main: 9.8% CPU / about 493 MiB RSS
• Athena process tree: about 2.5 GiB RSS with four live agent panes

This PR fixes those causes as one resource/lifecycle contract and adds permanent tests for the regressions seen across #26, #57, #95-#100, #146, #154, #169-#172, #181, and #184.

What changed

Terminal delivery, visibility, and layout

• Replace preload-dispatch ACKs with epoch/sequence delivery and ACK only after the xterm write callback drains.
• Retain the exact in-flight payload for retry; stale, duplicate, and wrong-epoch ACKs are harmless.
• Make attach atomic: a versioned bounded snapshot precedes contiguous live output for renderer and control/SSE consumers.
• Keep queues bounded per consumer. Overflow becomes an explicit reset snapshot rather than a silent wedge or unbounded heap.
• Store terminal history in bounded chunks and choose Unicode/ANSI-safe replay boundaries without repeatedly joining the discarded prefix.
• Cap renderer remount replay at 64 KiB while preserving the independent 200k control buffer.
• Stop broadcasting raw bytes to hidden panes; attention classification now emits small main-process state events.
• Separate visible from focused. Only the selected terminal owns focus, and one coalesced lifecycle/layout scheduler handles fit/refresh.
• Replace browser-owned pane heights with React-owned sizing and remove the press theme's full-window turbulence filter.

Session discovery and allocation bounds

• Move Hermes file parsing out of Electron main into a below-normal-priority metadata worker.
• Persist only compact metadata/path hints, not up to 300k characters of transcript search text per session.
• Make persisted-index load single-flight and retire/restart a wedged worker after timeout/send failure instead of reusing it forever.
• Bound Electron's workspace TTL/LRU cache and backend provider scans/caches; preserve last-known-good Hermes metadata on transient corruption.
• Apply workspace predicates before provider limits and align POSIX-case/root, Windows, WSL, and UNC identity rules across renderer, Electron, and backend.

Launch admission, restore, and process lifecycle

• Put UI, control, and restore terminal launches behind one atomic Electron admission service using physical MemAvailable only.
• Reserve 640 MiB per heavyweight agent tree, stagger every non-shell grid launch by 750 ms, and make critical-memory UI approval one-shot.
• Add equivalent bounded admission to the legacy FastAPI spawn route; return structured rejection/defer results instead of false success.
• Recheck pressure between restore entries, preserve missing provider identities instead of silently starting fresh, and keep restore crash evidence unless PTY shutdown is confirmed before the quit deadline.
• Bound backend stdout/stderr capture, use process groups, gate late spawns during shutdown, await owned service cleanup, and force-reap stubborn descendants.

Adaptive Linux graphics

• Healthy Linux launches use hardware acceleration, addressing the measured software-compositor heat.
• Persist an accelerated-pending marker before Chromium starts. An unclean accelerated exit or GPU-process crash quarantines the next launch into safe mode.
• Keep explicit safe/accelerated settings and environment overrides.
• Ensure a rejected packaged second instance cannot mutate the primary instance's graphics marker or log.
• Keep the Linux native beep disabled; later Harden terminal rendering and disable Linux native beep crash path #100 evidence implicated that path in the original crash sequence.

Regression protection and diagnostics

• Add permanent PR-lineage suites for terminal stream/ACK, GPU policy, pane layout/focus, session indexing, admission, and backend output bounds.
• Add CI for the full Python, renderer, Electron, production-build, and permanent-regression matrix.
• Surface terminal produced/delivered/acked/reset/replay counters, hidden raw IPC, session-index scan stats, event-loop lag, and process-tree diagnostics in Settings.

Why this avoids the previous regressions

• It does not restore Keep workspace terminals mounted and show quiet attention badges #57's keep-every-xterm-mounted design.
• It preserves Bound terminal output flow across Electron and xterm #169's bounded queues while fixing the lost-ACK wedge introduced by its one-batch gate.
• It does not land Bound terminal replay on attach #185's smaller arbitrary replay tail as a standalone fix; truncation is explicit and VT/Unicode boundaries are guarded.
• It folds Refit embedded terminals when windows regain focus #187's window-return behavior into one lifecycle manager instead of adding listeners/timers per pane.
• It implements Reflow terminal panes when a sibling is spawned #189's reflow with React-owned state and focus/layout regression coverage rather than browser inline-height mutation alone.
• It retains Athena Code session discovery + resume sessions across app restarts #141's restore identity/idempotency protections and does not let a failed shutdown erase its crash-loop signature.
• It replaces Harden terminal rendering and disable Linux native beep crash path #100's all-Linux GPU penalty with persistent crash quarantine instead of blindly enabling acceleration after a failure.

Validation

• Production client build: pass
• Renderer/chat/layout/path tests: 31/31 pass
• Electron/helper tests: 222/222 pass
• Permanent historical-regression tests: 72/72 pass
• Python/backend tests: 207/207 pass
• git diff --check: pass

Short hardware canary on the affected Surface Laptop 3 (Intel Iris Plus G7, 8086:8a52, i915):

• accelerated launch and clean window shutdown completed without a runtime GPU-crash signal;
• one mounted xterm processed a bounded 5,000-line shell burst;
• after the burst settled, the renderer and GPU processes averaged about 1% combined CPU over a five-second sample;
• graphics and launch state both recorded a clean exit.

This canary validates startup, mounted-xterm rendering, and clean fallback state on the known-risk device. It is not a substitute for the long soak below.

Known limits / follow-up gates

• No two-hour CPU/RSS soak or sustained four-pane 64 KiB/s hardware load was run in this PR session.
• No real Windows ConPTY hardware run was available.
• Bounded reset/replay is explicit and safe, but it is not yet a full authoritative serialized VT checkpoint; sufficiently old scrollback may be intentionally discarded.
• Electron and legacy FastAPI admission reservations are process-local. A narrowly concurrent legacy HTTP spawn and desktop spawn can still race before OS RSS becomes visible; the legacy path is separately capped and was not implicated in the measured incident.
• External CLIs that deliberately self-daemonize outside their inherited process group still need release-soak verification.

Checklist

• Reviewed the prior performance/freeze PR lineage and reconstructed regression chains
• Added regression tests before changing the riskiest contracts
• Preserved bounded-memory and restore/idempotency invariants
• Exercised accelerated Chromium and a mounted xterm on the affected Surface hardware
• Added rollback controls for Linux graphics mode
• Complete the two-hour multi-pane soak and Windows ConPTY hardware gate before release promotion

[luckeyfaraday]

Manual validation update: the resource-stability AppImage built from a497aa4 was tested by the user on the affected machine; they report that it looks good and works. This is additional smoke-test evidence. The two-hour multi-pane soak and Windows ConPTY hardware gate listed in the PR remain open.