revert(tui_gateway): drop the heavy session-read admission gate (#215, #441)

[Kyzcreig]

Fork-PR audit UNRESOLVED ruling: DROP #215 + #441 (the re-land), reverted as ONE unit. Card t_7c17872a (from t_63023f77; rulings doc PR #1186).

What gets removed:

• hermes_cli/session_db_heavy_gate.py (deleted)
• tui_gateway/ws.py: the gate import and _dispatch_ws_request. WS requests go back to asyncio.to_thread(server.dispatch, req, transport).
• tui_gateway/server.py: _SESSION_DB_HEAVY_METHODS, is_session_db_heavy_method, backend_busy_error, handle_request_bound. The data= parameter on _err stays because methods_prompt.py and methods_bot_relay.py now use it.
• hermes_cli/web_server.py: the gate import, the 503 shed handler and the /api/status session_db_heavy_reads field. _session_db_read(heavy=True) goes back to the per-loop asyncio.Semaphore(2) it used before feat(tui_gateway): async-side shared heavy-read bound for session.list (seam a) #215.
• config_defaults: dashboard.heavy_read_max_concurrency
• The gate tests in tests/test_web_server_sessiondb_eventloop.py and tests/scripts/test_preyield_permit_release.py, plus the live-site line in scripts/check_preyield_permit_release.py. The guard sweep still has one live site (gateway/turn_admission.py::TurnAdmission.slot), and that site's named coverage test stays, so the guard is not vacuous.

A plain git revert conflicted (UD on the gate module; UU on config.py, server.py and ws.py). The defaults had moved to config_defaults.py, and the checkout-admission and event_replay code grew around the same hunks. All of these were resolved by hand.

Why drop it: 0 session_db_heavy_read log lines across ~/.hermes/logs and profiles/*/logs (the gate logs any wait of 1 ms or more), so it has never contended under fleet load. It is 473/474 loc across 5 files, and those files conflicted in all 3 upstream syncs. Upstream bounds reads in its SessionDB read pool instead. One caveat: the live /api/status counter is auth-gated and was not read.

Verification:

• test-gate (ACE-AI), running pytest on tests/test_web_server_sessiondb_eventloop.py tests/scripts/test_preyield_permit_release.py tests/test_tui_gateway_ws.py tests/tui_gateway/test_ws_keepalive.py: 85 passed
• ruff --select F on the touched files: no new findings compared with origin/main
• git grep finds no remaining reference outside docs/ to session_db_heavy_gate, SessionDBHeavyReadBusy, session_db_heavy_reads, heavy_read_max_concurrency, handle_request_bound, backend_busy_error or _dispatch_ws_request

Not merged from the worker lane; open for the merge pass.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[ang-fleet-lander]

🤖 merged-by: daedalus-opus · lane: t_f201acde · gate: BYPASS: FleetReview paused by Ace 2026-09-22 (state/fleetreview-pause marker present) · why: fork-PR audit DROP (Ace ruling 1A, campaign t_03e35f0e, card t_7c17872a): revert-#215+#441; PR CI green on head; conflict-free vs fork/main + sibling reverts (merge-tree chain)

[Kyzcreig]

Superseded by batched PR #1230 (card t_7c5e08a5). It carries this revert unchanged: the batch merges this PR's head, and every touched file is byte-identical or hunk-identical. Closing so the merge queue carries 1 entry per audit area instead of 1 per revert. The branch is kept.