feat(processor): add model fallback chain when retries are exhausted - #42424

Agents can now configure a fallback model chain (agents.<name>.fallback
in config, or the older top-level agent: format). When the selected
model fails after RequestExecutor's transport retry budget is spent
(429/5xx/529), or with a quota/transport/model-not-found failure, the
V2 session runner switches to the next untried fallback and retries
the turn -- rather than erroring the whole turn out to the user.

Ported from real, working upstream implementations rather than written
from scratch (per the standing "reuse upstream PR work" directive for
this fork): anomalyco#35188 already targeted this repo's V2
session runner (unlike anomalyco#49125/anomalyco#26292/anomalyco#27939/anomalyco#42424, which patch the
legacy V1 processor.ts our own AGENTS.md says new orchestration code
must not bridge through) and is the primary source for the config
schema shape and general approach. The retry-exhaustion classifier's
names/shape come from anomalyco#49125; its logic reuses RequestExecutor's
existing LLMError.retryable flag as the single source of truth for
"is this retryable" instead of re-deriving a second status-code list.
Model-switch attribution is recorded via the session's existing
SessionEvent.ModelSwitched event rather than mutating an in-flight
input object in place -- a lesson taken directly from anomalyco#27939's
writeup of a real bug in its own V1-era approach.

New packages/core/src/session/runner/fallback.ts (~40 lines: a
classifier + a candidate resolver) and a new ContinueAfterFallback
turn-transition in packages/core/src/session/runner/llm.ts, following
the exact pattern already used for post-compaction turn continuation.
Config plumbing threads the new field through packages/schema/src/
agent.ts, packages/core/src/config/agent.ts (+ plugin and v1 config
variants). Regenerated the client and legacy JS SDKs.

Known, honestly-scoped gaps: the switch is sticky for the rest of the
session (no cooldown/return-to-primary); no mid-stream failover once
assistant output has started; only errors marked retryable trigger a
switch, so some in-stream provider errors won't while an equivalent
HTTP status will; fallback is per-agent only, not per-model or
global; a fallback model uses its default variant; no TUI/docs yet.

New packages/core/test/session-runner-fallback.test.ts (9 tests)
exercises real RequestExecutor retry counts against a scripted fake
HTTP client -- exact call sequences, session projection correctness
(model-switched message inserted, no leaked failed assistant message,
correct final session model), chain-walking past an unresolvable
model, non-retryable failures correctly not switching, and chain
exhaustion still surfacing the real terminal error. Full packages/core
suite: 1164 pass (only the pre-existing, unrelated pty flake). bun
typecheck clean across all 30 packages.

Co-Authored-By: Claude Sonnet 5 <[email protected]>