feat: Bedrock Converse codec and typed routed-call failure contract

[c1-squire-dev]

Summary

Two independent public-surface additions staged on this fork branch for review.

Rebased onto current upstream main (5879c10). This branch previously sat on
fb3fc30, 133 commits behind. main on this fork was a stale mirror with no commits of its
own and has been fast-forwarded to the same revision, so this PR is exactly the three commits
below.

Upstream had moved the code these changes touch, so the second and third commits were ported
rather than replayed: routing fallback classification now lives in
libsy-llm-client::run::fallback_reason (it moved out of libsy::core::algorithm), the
escalation classifier lives in libsy/src/algorithms/escalation.rs, and the new
switchyard-runner::failure classifier is an additional dispatch point that did not exist
before. Per-session context-window eviction memory was deleted upstream; see the third commit
for how that affects the overflow case.

Bedrock Converse translation codec

• add bedrock_converse as a first-class protocol format
• translate buffered Bedrock Converse requests and responses through Switchyard’s neutral model
• decode and encode AWS EventStream semantic events incrementally
• cover tools, reasoning, guardrails, usage, stop reasons, malformed events, and cross-format round trips

Typed routed-call failure contract

Lets a routing host classify its own call failures instead of having libsy re-derive
fallback policy from an HTTP status.

• add RoutedCallFailure, RoutingDisposition, and the non-exhaustive
  RoutedCallFailureClass with immutable stable_tag() values to switchyard-protocol,
  plus LlmClientError::RoutedCall
• add bounded provider-exhaustion evidence (RoutedFailureCount,
  ProviderTargetsExhaustedSummary). Validating constructors are the only way to build
  these, so a value always satisfies the contract: candidates bounded to 16, unique
  nonzero class entries that partition attempts and circuit-open bypasses, no nested
  exhaustion, and retry advice capped at 300000 ms
• libsy fallback now follows the host's explicit RoutingDisposition for a RoutedCall.
  The existing 403/408/429/5xx inference is untouched for the transport- and
  status-shaped variants, so current clients keep their behavior
• map RoutedCall at every semantic dispatch point: libsy fallback classification, judge
  bounded reasons, client observability labels (reusing the immutable stable tag), and
  switchyard-server status/body mapping with a rounded-up Retry-After when the host
  supplied truthful advice

The types are provider-neutral by construction: bounded class tags, an optional bounded
provider status, and counts only — no provider body, message, or target name.

Follow-up commit fix(routing): honor routed-call classes at every fall-through and status seam closes three gaps found in review of the commit above:

• the staged-escalation fall-through matched only the legacy ContextWindowExceeded, so a
  host returning RoutedCall terminated the run instead of escalating to the capable tier.
  It now also accepts an advanceable ContextWindow or TargetIncompatible, and still
  treats Stop as terminal
• ProviderTargetsExhausted collapsed to 503 with its bounded summary unread. The status
  now resolves from the partition in precedence order — exact availability evidence or a
  circuit bypass, then host or provider timeouts, then the homogeneous rate-limit and
  request-shape cases, then a plain routing failure — so the 429, 504, and 422 codes are
  reachable and 503 stays truthful. No provider status is consulted
• an exhaustion whose every real failure was an overflow reported target unavailability,
  which is exactly the conflation this contract exists to remove. It is now reported as a
  context-window condition at both seams that consume the distinction: the candidate hop's
  reason, and the runner's terminal telemetry kind, where it previously fell into Other.
  Bypassed circuits neither establish nor contradict the overflow, and an all-bypassed
  exhaustion stays unavailability; mixed classes remain unavailability

ProviderTargetsExhaustedSummary grows real_failures and is_context_window_exhaustion so
the aggregate rules are stated once for all three consumers, and rejects an over-long
class-entry list before it sizes an allocation. No existing public type, constructor
signature, or client requirement changes.

Note for reviewers: upstream deleted per-session overflow memory, so the repeat-overflow cost
the third bullet originally eliminated is real again on this base for every client, not just
routed-call hosts. This branch keeps the classification truthful but does not re-add the
memory — that reverses an upstream decision and is worth deciding deliberately.

Validation

• cargo fmt --all --check
• cargo build --workspace
• cargo test --workspace -- --test-threads=1
• cargo clippy --workspace --all-targets -- -D warnings
• RUSTDOCFLAGS=-D warnings cargo doc --locked --no-deps --package switchyard-libsy --package switchyard-protocol
• translation conformance suites for buffered and streaming requests/responses

This fork PR is the review staging point before a human submits the change upstream to
NVIDIA-NeMo/Switchyard, per D-TB-103.