fix(tui_gateway): close dedicated profile SessionDB handles on every path (salvage #76701)

[kshitijk4poor]

Summary

Dedicated profile-scoped SessionDB handles are now closed on every path — the pre-transfer early returns in session.resume, the deferred/branch/compute-host build failure paths, the reaped-mid-build discard, and agent teardown (AIAgent.close() releases a handle it owns). Fixes the fd leak class where long-lived Desktop/serve backends accumulated 2 fds + a pinned token-writer thread per abandoned handle until OSError: [Errno 24] Too many open files.

Salvage of #76701 by @Yishova — both commits cherry-picked with authorship preserved. Supersedes #80706 (strict subset: teardown half only, tests break on macOS), overlaps #78970/#72804 at other layers.

Who hits this

Anyone running the Desktop app or hermes serve against multiple profiles for days: every reconnect/tile-paint resume of a live chat (session.resume fast path — the hot one), every default cold resume, and every successfully-built-then-closed agent leaked a SQLite handle. An abandoned SessionDB is never GC'd once its background token writer starts — it pins ITSELF via atexit.register(_drain_token_queue_at_exit), which only close() unregisters. Observed live: ~490 leaked fds over ~2 days, backend wedged at the 1024-fd soft limit.

Changes

• tui_gateway/methods_session.py — session.resume gets explicit owns_db ownership with a try/finally close covering all ten pre-transfer early returns; ownership transfers exactly at _init_session success; half-built session registration rolled back on failure (prevents a permanently-dead live-session fast path).
• agent/agent_init.py + run_agent.py — _owns_session_db lifecycle flag (mirrors _end_session_on_close); the recall fallback's self-created handle is owned; AIAgent.close() releases an owned handle after end_session(), clearing the flag first for idempotency.
• tui_gateway/server.py + compute_host.py — deferred builder, branch handler, and compute-host paths transfer ownership via _transfer_db_to_agent() (refuses mismatched handles; closes directly when the built agent was discarded mid-build).
• Tests: 654 lines across 2 files, platform-independent _RecordingDB doubles — owned-closed-exactly-once, shared-never-closed, idempotency, transfer refusal, reaped-mid-build, registration rollback, end_session → close ordering.

Note the raw diff on methods_session.py is mostly re-indentation from the try: wrap; review with git diff -w (~40 insertions / 2 deletions there).

Validation

Check	Result
Ownership test files + tests/tui_gateway/	373 passed (+1 new raising-close contract test → 22 in the ownership files)
tests/test_tui_gateway_server.py	519 passed
Mutation check (revert the 4 production files to upstream/main)	16 ownership guards fail on base, green on stack
Integration review of the re-applied commit-1 hunk	every return in session.resume covered by the finally; try starts after db binds; main's churn (workspace/steer/omit_messages) all inside the try; no untracked opens in tui_gateway/
Cherry-pick conflict resolution	commit 1 re-applied onto current main's session.resume (whitespace-normalized hunk set verified identical to the original)
ruff	clean

Related PRs in this cluster

• fix(run_agent): release owned SessionDB connection on agent close #80706 — same _owns_session_db teardown mechanism, opened 5 days later; subset of this PR (misses all pre-transfer paths, marks the reaped-mid-build agent owner without anything closing it, /proc/self/fd tests hard-fail on macOS). Recommend closing with credit once this merges.
• fix(tui): close profile-owned SessionDB handles on teardown #78970 — gateway-layer close in _teardown_session only; fix: close SessionDB FDs on timeout and lazy recall paths (#72782) #72804 — timeout + lazy-recall paths only; fix(state): close threaded SessionDB read conns without leaking FDs #74304/fix(hermes_state): bound per-thread read-connection cache to prevent fd leak #75546/state: pool SessionDB read connections instead of leaking one per (SessionDB x thread) #76700 — per-thread READ connection cache (different leak, separate fix).

Based on #76701 by @Yishova — commits cherry-picked to preserve authorship.