fix(tui_gateway): isolated turns no longer fence themselves out of their own session lease (#101416, salvage #103737)

[teknium1]

With dashboard.turn_isolation: true, a new Desktop session's first message no longer fails with "This chat is open in another Hermes window/terminal" — the compute-host child now runs under the parent's lease, and that lease stays singular through compression rotation and through session.close.

Fixes the compute-host variant of #101416 (the child fenced out by its own parent's registry entry). VIPKaiser's cross-process route-split variant in that thread is a different bug and is handled by another lane, so this PR does not close the issue.

Changes

• Salvage of fix(tui_gateway): stop isolated turns fencing out the parent's session lease #103737 (@jakobbjelver, authorship preserved): the parent vouches on the turn.start frame and the child installs an inert ActiveSessionLease(enabled=False) before _run_prompt_submit, so _admit_prompt_turn sees the slot as held upstream instead of re-claiming from the child pid (session_lifecycle._install_borrowed_lease, compute_host.ComputeHost._run_real_turn).
• Review blocker 2 (@andrexibiza) — rotation A→B split authority: the borrow is no longer released=True (a released token made transfer_active_session return False and the fallback claimed a REAL lease for B under the child pid). The vouch is qualified — {lease_id, session_id} — and the child only borrows when it names the admitted stored id; the parent re-anchors its real lease A→B where it adopts the child's rotated session_key (compute_host_bridge._compute_host_adopt_frame_meta, on turn.end and compress acks). A stale A lease never vouches for B (compute_host_bridge._active_session_lease_vouch).
• Review blocker 1 — close released the lease while the child was still writing: session_lifecycle._teardown_popped_session now interrupts the isolated turn, waits the same close grace, and if the child has not settled moves the real lease out of finalize's reach; the correlated turn.end/turn.error (child death fires it via _fail_pending_turns) releases it (_settle_isolated_turn_before_close, _release_deferred_active_session_lease). Deferred leases stay live authority for the orphan sweep (_own_live_lease_ids).
• Tests trimmed from 383 lines / 10 tests to 4 invariants (one per fix + the fail-closed negative control); 3 red on origin/main, 4 green here.

Root cause

The parent claims the registry lease in prompt.submit before routing, but the child's freshly built session record had no lease, so it re-claimed from its own pid and _is_same_writer (same pid AND same live id) refused it.

Live evidence (real compute-host child, temp HERMES_HOME, HERMES_ISO_CERTIFY_SYNTH_TURN=1)

Probe	origin/main @ bbaf7af	this head
First isolated turn	child: Refused active session stored-A: already held by pid=<parent> surface=desktop → client gets This chat is open in another Hermes window/terminal… on every attempt	message.complete '[synthetic heavy turn] …'; registry = exactly one entry {pid: <parent>, session_id: stored-A} across 2 turns
A→B rotation inside the child (B2)	(pick-only) registry = [{parent, stored-A}, {child, stored-B}]; replacement child ran under B on the stale A vouch	registry = [{parent, stored-B}] after rotation, after a second turn, and after a child restart; parent lease.session_id=stored-B
session.close while child turn live (B1)	close returned in 0.5s, registry [], foreign try_acquire_active_session succeeded mid-turn	close returned after the 5s grace with holding lease for stored-A until the child settles; foreign acquire refused (SESSION_NOT_OWNED); after turn.end registry [] and foreign acquire succeeds
Control: turn_isolation: false	in-process turn completes, one parent lease	unchanged

Suite: scripts/run_tests.sh tests/tui_gateway tests/hermes_cli → 14020 passed, 5 failed — test_dashboard_auth_gate (4, also red on bare origin/main on this host) and one test_profiles_sidebar_cache / test_auth_commands timing flake under 40 workers that passes in isolation; none touch this PR's files.

Not covered

• An interrupt frame does not stop an isolated synthetic turn in the child (session["running"] reads False in the child mid-turn, so _interrupt_session_turn no-ops there) — pre-existing on origin/main, orthogonal to lease ownership; the deferred-release path is what guarantees ownership here.
• Mid-turn compression rotates agent.session_id before the post-turn _sync_session_key_after_compress; the window until turn.end is the same one the in-process path has today.

Credits

• @jakobbjelver — diagnosis and the borrowed-lease carrier (fix(tui_gateway): stop isolated turns fencing out the parent's session lease #103737, cherry-picked).
• @Bergmann89 — deferred canonical-lease lifetime and owner-side re-anchor design (Fix/dashboard turn isolation profile scope #101501, Co-authored-by on the two follow-up commits).
• @andrexibiza — the two blockers this closes.
• Reporters: @rmprojetotrader-ops, @VIPKaiser.

Infographic

Independent review (pre-merge) — dispositions

• Attribution check red: jakobbjelver@gmail.com unmapped → mapping file added (1c60aab). The inherited Closes #101416 trailer was reworded to a Related reference (tree byte-identical) so this PR does not close the issue's route-split half (Desktop: pooled-profile chats no longer land on the local primary and get refused SESSION_NOT_OWNED after reload (#101416 class) #113956).
• Known limitation (kept): a compute-host child that never settles after session.close holds the deferred lease until the parent exits (no supervisor stall-kill exists today); a same-window re-open during that window is refused with the "open in another Hermes window" copy. Bounded second grace + force-terminate is a follow-up, not folded in here.
• Edge (same as main today): if the parent-side A→B transfer fails for a liveness-tracked lease, session_key advances while the lease stays on A; the next turn has no vouch and takes the legacy self-claim path (fail-closed refusal, never a silent second writer).
• Reviewer verified: SIGKILL of the child → _fail_pending_turns → _on_compute_host_turn_done → lease released (registry empty, foreign acquire succeeds); no double release on the ws-orphan path; 806 lifecycle/orphan tests + 71 sibling tests green on head.

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 1c60aab — chore: map contributor email for @jakobbjelver}

debug info

CI timings

CI timings · View report · View job

Wall time 8m40s vs 46m37s (-81.4%). 5 job(s) slower, 6 faster, 1 unchanged.

• Python tests / Run tests: +153.0s
• Python lints / Windows footguns (blocking): -52.0s
• OS-specific tests / Windows-only tests: +20.0s
• OS-specific tests / macOS-only tests: +11.0s
• Python lints / ruff enforcement (blocking): -8.0s

[alt-glitch]

This was generated by AI during triage.

Related: #103737 is the earlier parent-vouch/borrowed-lease attempt for #101416. This maintainer salvage retains that direction while adding qualified lease/session vouching, rotation re-anchoring, and close-settlement handling.