revert(failover): drop Tailscale transport-outage guard (#621, #622) — fork-PR audit DROP

[Kyzcreig]

Fork-PR audit DROP (lead t_03e35f0e, FINAL.md rows #621/#622; card t_e5ecb888).

Reverts #622 (keep Tailscale errors terse) and #621 (bound Tailscale transport outages). Auditor verdict DROP: 'Tailscale' appears in 1 line across 54 rotated gateway/agent logs; the adversary lens let it stand.

Two commits, rebased onto main 382fc62:

• 88d1fa3 Revert fix(failover): keep Tailscale errors terse #622 (clean).
• 13ec95a Revert fix(failover): bound shared Tailscale transport outages #621, hand-resolved:
  • agent/chat_completion_helpers.py, agent/conversation_loop.py: Tailscale preflight blocks removed. _POOLED_PROVIDERS = frozenset({claude-apr, claude-bpr}) is inlined because feat(blackbox): C2 accumulator — every upstream call appends a turn_api_calls row with wire/pinned subscription attribution #913 imported _TAILSCALE_RELAY_PROVIDERS from the deleted module.
  • agent/turn_context.py (rebase conflict): the shared_transport_guard.reset_turn_state call is removed and the quota_registry_gate.reset_quota_gate_turn_state reset that landed later on main is kept.
  • agent/chat_completion_helpers.py, dedupe hunk NOT reverted (KEEP-by-dependency): fix(failover): bound shared Tailscale transport outages #621 also removed the buffered "🔄 Primary model failed — switching to fallback" line, which duplicated each hop. That change has nothing to do with Tailscale, and feat(turnover): skip registry-known-exhausted subs, and leave a durable handoff when a turn is cut #813's tests/agent/test_quota_gate_fallback_wiring.py (the switching-spam collapse tests) depend on it. The first push of this branch reverted it and CI slice 13 went red with 2 failures; main's version is kept now.
  • tests/agent/test_error_classifier.py: tailscale_down removed from the enum list; pool_stalled kept.

Deleted: agent/shared_transport_guard.py, tests/agent/test_shared_transport_guard.py, tests/run_agent/test_shared_transport_fallback.py. Nothing on the branch still references shared_transport_guard, _TAILSCALE_RELAY_PROVIDERS or tailscale_down.

Registry (docs/sync/fork-features.json): no entry lists a deleted path or mentions Tailscale. Entry 25 (Claude CLI usage-cap classification) shares three of its paths with this diff, but none of its anchor lines are in the diff: _CLAUDE_CLI_USAGE_CAP_RE, usage exhausted, quota_window_reset_text and quota_window.

Local narrow runs through test-gate:

• test_claude_cli_usage_cap_classification + test_error_classifier + test_fallback_reason_threading: 165 passed.
• test_quota_gate_fallback_wiring + test_fallback_reason_threading + test_fallback_reason_surfacing: 44 passed. Before the dedupe hunk was kept, test_quota_gate_fallback_wiring had 2 failures on this branch and passed 14/14 on main.
• test_fallback_reason_surfacing + test_user_providers_model_switch: 29 passed, 2 failed. Both failures are inherited: the same 2 node ids also fail on clean main 239f717 (test_list_authenticated_providers_enumerates_dict_format_models, test_section3_probes_no_key_endpoint_with_singular_default_model).

[ang-fleet-lander]

🤖 merged-by: apollo · lane: kanban-merge-pass · gate: BYPASS: FleetReview paused by Ace 2026-09-22 (state/fleetreview-pause marker present) · why: t_e5ecb888: fork-PR audit DROP: #621 fix(failover): bound Tailscale transport outages (+#622; Argus off card review (Ace 13:08), CI green

[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_e5ecb888): revert-#621+#622; PR CI green on head; conflict-free vs fork/main + sibling reverts (merge-tree chain)

[ang-prism]

FleetReview

PARTIAL — ensemble escalated: family floor: too few distinct model families completed

This review did not reach a trusted verdict, so it is not a gate pass and the findings below may be incomplete. They are posted so they can be read rather than lost in a terminal record.

Post-merge review (fleetreview:post-merge override): this reviewed the merge commit against its first parent — the bytes that already shipped. It is not a pre-merge gate pass.

Reviewed with 1 of 3 model families — anthropic, openai unavailable.

Confidence: 1/5

Findings

• P1 agent/conversation_loop.py:3632 — Tailscale fail-fast removed
• P1 agent/chat_completion_helpers.py:3771 — Dead-route fallback
• P1 agent/shared_transport_guard.py:1 — Guard Removed
• P1 hermes_cli/model_switch.py:2397 — Tailscale preflight removed
• P2 tests/run_agent/test_shared_transport_fallback.py:1 — Tests Removed
• P1 agent/chat_completion_helpers.py:71 — Pool set mismatch

---

FleetReview provenance · models: C=claude-code-opus-5-5, D=grok-4.6 · cost: $0.00 · duration: 13m 11s · rounds: 3 · files examined: 12