fix(escalation): require fresh consistent failure evidence

[pst2154]

What

• Normalize duplicated Terminus command serialization in the escalation judge's transcript without changing model-facing history.
• Require consecutive escalation verdicts to cite fresh evidence for the same bounded failure category before latching to the strong tier.
• Log the parsed category, freshness flag, and a bounded reason; update the packaged schema, prompt, tests, and routing documentation.

Why

Agent harnesses can represent one command both as assistant JSON/text and as a structured tool call. The trajectory judge could count those same-turn representations as repeated failed attempts. More generally, boolean-only confirmations could combine stale evidence or unrelated failure modes into a false escalation.

The new confirmation rule keeps the existing one-way latch while requiring a coherent multi-turn failure pattern. This is a focused correctness improvement related to the signal-composition direction discussed in #601.

Live SWE-Atlas diagnostics with GLM 5.2 as the efficient tier and Opus 4.8 as the capable tier:

• Positive path, RF task-697e7458be1623d850a88838: two consecutive fresh repetition verdicts caused exactly one GLM-to-Opus transition; the run used 28 GLM and 36 Opus worker calls and received verifier reward 1.
• Recovery control on the same task: two repetition votes separated by a decline did not latch; the GLM-only run received verifier reward 1.
• Prior false-positive cases RF task-69391d8d1ce51c407be1e531, RF task-694b4b99829f00e24fd11889, and TW task-6902ef3ab97fe23e2ad271f3 all remained GLM-only under the new rule. In the first case, positive votes changed from repetition to false_progress, so they did not combine into a latch.

These are routing diagnostics, not an aggregate accuracy claim: the positive-path task also passed in a separate direct-GLM rollout, and the false-positive controls still received reward 0. The evidence shows correct confirmation/latch behavior and avoided strong-tier calls on previously wasteful trajectories.

Notes for reviewers

Start with crates/libsy/src/algorithms/escalation.rs for the confirmation state machine and crates/libsy/src/algorithms/util/escalation.rs for judge transcript normalization and typed verdicts.

Validation on Rust 1.96.1:

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --all-features
• uv run ruff check .
• uv run mypy switchyard
• wheel build/install through maturin
• uv run pytest tests/ -v (117 passed, 2 subtests passed)

[pst2154]

Superseded by #637, which carries the current fresh-evidence implementation and paired SWE-Atlas report. The separate directional/model-order follow-up is #638.