Conversation momentum implementation

[QuMuzafferEge]

Summary

Adds an opt-in routing.momentum policy block to the router engine to fix route flapping in multi-turn conversations. Today min_chars/max_chars only ever measure the latest user turn, so a long coding question correctly escalates to a strong model via min_chars, but the very next short follow-up ("yes, do it") falls straight back to default_model mid-task — paying a real switch cost (model load/eviction, KV re-prefill, provider cache miss) every time.

routing.momentum replaces the raw last-turn byte length with a momentum-filtered "effective length," recomputed fresh from each request's own message history (no server-side session state — the conversation is already resent every turn, same as any OpenAI-style chat client):

json
"routing": {
"momentum": { "enabled": true, "attack": 1.0, "release": 0.3 }
}

An asymmetric EMA over each user turn's UTF-8 byte length: attack (default 1.0) weights a rising turn so a sudden long/complex turn escalates immediately, release (default 0.3) weights a falling turn so a short follow-up only decays the effective length by 30% per turn — keeping the conversation on the escalated route for several turns before de-escalating. attack = release = 1.0 degenerates exactly to today's behavior. Applies only to min_chars/max_chars; absent or enabled: false is behavior-identical to a policy without the block.

Fixes #2956

Scope
This PR addresses one clear issue or change.
I reviewed the full diff myself before submitting.
I removed unrelated local changes.

Testing
Code builds without errors locally.
I tested this change locally.
I described the testing performed below.

Testing details:

Built and tested on Linux (Ubuntu 24.04 container, matching the cpp-unit-tests CI job's recipe: cmake --preset default → cmake --build --preset default --target cpp-ci-tests → ctest -L "^cpp-ci$").

50/52 cpp-ci tests passed, including every test this change touches:
RoutingPolicyMomentumTest (new) — filter arithmetic (rise/fall/tie/degenerate coefficients, IEEE-exactness), CharsCondition/EvalContext::effective_chars wiring, engine-level fold, trace rationale format.
RoutingPolicyParserTest (extended) — momentum block validation (required enabled, (0,1] coefficient bounds, unknown-key rejection) and schema/parser key parity.
RoutingClassifierServicesTest (extended) — build_route_context collecting user_turn_chars across chat messages, legacy prompt, and Responses input forms.
RoutingConformanceCorpusTest — two new golden groups (l1_momentum_effective_chars, l1_momentum_absent_is_noop) plus every pre-existing group re-validated byte-for-byte unchanged, confirming no regression when the block is absent.
2 pre-existing failures (TelemetryHelpersTest, CliRuntimeOverrideTest) are unrelated to this change — neither file is touched by this diff, and both look like environment artifacts of a from-scratch container vs. the pinned CI image (locale/encoding and a streaming-timing sensitivity respectively).
python test/test_schema_lock.py and python -m unittest test.test_routing_fixtures both pass (schema-lock hash refreshed for the additive routing.momentum schema keys and the decision.schema.json rationale-field description update).

Documentation
Documentation is affected and has been updated.

Added a "Momentum (routing.momentum)" section to docs/dev/router-policy.md (field table, formula, worked example, trace format, explicit non-goals), footnoted the min_chars/max_chars schema descriptions and the frozen-semantics table in src/cpp/resources/schemas/README.md to reference the new opt-in mechanism without redefining the existing frozen unit/comparator.

Breaking Changes
This PR does not introduce breaking changes.

Fully additive and opt-in: new RoutePolicy/RouteContext/EvalContext fields (C++-internal), one new optional routing.momentum schema property, and use of the already-optional trace_entry.rationale field. No existing field is redefined, and every code path is gated behind momentum.enabled (default false), so any policy without the block is behavior-identical to before this change.

[jeremyfowers]

Please give this PR a descriptive title before merging.