refactor(agent): extract fallback policy into auxiliary_fallback_policy (auxiliary_client.py R3-S1)

[andrexibiza]

Summary

agent/auxiliary_client.py god-file slice R3-S1: extract the runtime fallback policy cluster (window 4891–5369, 479 lines, 9 defs) into agent/auxiliary_fallback_policy.py. Part of the repo-wide large-file decomposition (tracker #78647, target #78635).

What changed and why

• 9 defs moved byte-verbatim (golden sha256 c365b2fde16264d31cb79e577ad4519162ac172aea689fa273aca215ef7ded11 verified; zero unsanctioned byte differences)
• Moved: _try_payment_fallback, _try_main_agent_model_fallback, _task_minimum_context_length, _candidate_context_window, _try_configured_fallback_chain, _try_configured_fallback_for_unavailable_client, _fallback_entry_api_key, _resolve_fallback_entry, _try_main_fallback_chain
• Re-export block at the old location keeps all 9 names resolvable as agent.auxiliary_client.<name> with __func__ identity (tests patch these names; agent/conversation_compression.py:1595 imports one of them)
• Outward edges (6 names) are function-local lazy imports — no import cycle, patch-through-monolith probes observed patched globals (no silent no-op seam)
• Window collision-free: consensus census verified zero open-PR hunks in 4891–5369; fix(copilot): support GitHub Enterprise endpoints #78378's hunk sits below the window (in _auth_refresh_provider_for_route)
• Double-blind: 2 blind analysts → R3-CONSENSUS → blind implementer → 2 independent blind witnesses (pass A + adversarial pass B) — both VERIFIED, zero failing checks

Testing

• 213 passed / 2 failed (4 gate files) in worktree; the 2 failures are byte-identical at pristine origin/main baseline (environment class, proven — not slice regressions)
• Seam identity: all 9 names is-identical in fresh interpreters, both import orders
• Standalone import agent.auxiliary_fallback_policy succeeds (no eager monolith import)
• py_compile clean · LF-only · git diff --check clean · DCO signed

Coordination / interlock

• Epic: Part of [COMPLETE] Large-file decomposition: 20/20 done #78647 · Target: Part of Shard agent/auxiliary_client.py (god-file decomposition) #78635
• Sibling PRs: auxiliary_client slices R4-S1 (anthropic-compat) and R5-S1 (text leaf) — disjoint windows, all three consensus-approved
• Dedup: no existing PR extracts this cluster (consensus census over 161 file-touching open PRs)
• Land-order: consensus execution order — fix(copilot): support GitHub Enterprise endpoints #78378 → Fix: Preserve anthropic_messages api_mode during fallback activation #79787 → fix(auxiliary): never send a model name to a foreign endpoint #79792 → R3-S1 → post-fix(copilot): support GitHub Enterprise endpoints #78378 W1 wave → clusters A+B → final router kill. R3-S1 itself is EXTRACTABLE-NOW (window hunk-free)

Part of #78647
Part of #78635

[andrexibiza]

Decision recorded against current main 3a034356a237341452b7afebd3a2bfd21021609f.

Do not land this fallback-policy slice first. The maintained auxiliary sequence is:

1. refactor(agent): extract anthropic-compat helpers into aux_anthropic_compat (auxiliary_client.py R4-S1) #80139 — Anthropic compatibility leaf
2. refactor(agent): extract extract_content_or_reasoning into auxiliary_text (auxiliary_client.py R5-S1) #80140 — text-normalization leaf
3. refactor(agent): extract fallback policy into auxiliary_fallback_policy (auxiliary_client.py R3-S1) #80138 — fallback policy

The historical #79787 gate is no longer pending: its work was salvaged to main through #85576. The active semantic gates are #79792, which still changes auto-resolution/provider-coherence behavior in this ownership area, and #78378, whose current Enterprise-endpoint disposition must be reconciled before fallback policy is frozen into a new module.

Required branch work: rebase after #80139/#80140, retarget any surviving #79792/#78378 logic to the canonical owner, fail closed on duplicate ownership, recompute the exact source-window/golden proof, and rerun current-main fallback/resolver suites. The decision is complete; the current head is not yet landable.