feat(escalation): gate switching on calibrated utility

[pst2154]

Summary

• add optional expected_capable_gain calibration to the escalation route
• bypass trajectory judging and serve the efficient target when the configured gain is nonpositive
• preserve existing behavior when the value is unset or positive, including normal transport/context fallback
• expose the setting through TOML and Python and document the deployment utility formula
• include a sanitized 100-cell SWE-Atlas TW reproducer and full findings report

This is an independent PR based directly on main. It complements the fresh-evidence fix in #637 but does not include that PR's commits.

Held-out evidence

Twenty deterministic held-out SWE-Atlas TW tasks were run once under each of five arms (100/100 final valid cells):

Arm	Correct	Synthetic total	Judge/classifier calls	Opus worker trajectories
Direct GLM	4/20	$6.422	0	0
Always Opus	7/20	$10.838	0	20
Patched escalation (#637)	7/20	$12.479	674	2
Capability selector	9/20	$5.036	20	0
Calibrated escalation (-0.10)	8/20	$7.520	0	0

The calibrated arm's routing invariant held across all 800 calls: every call was a GLM worker call, with no judge or Opus call. The uncalibrated patched arm had one long-tail task that made 146 judge calls and 144 GLM calls before a valid late switch, then 30 latched Opus calls; it scored 0 and cost $7.392 synthetically by itself.

The accuracy point estimates are not treated as causal. Three independently sampled GLM-worker arms scored 4/20, 8/20, and 9/20, and held-out Opus beat direct GLM despite the calibration sample showing the reverse. The report therefore concludes that this gate is a mechanical safety constraint when external calibration is reliable, not an automatic model-order solution. A full #601-style solution needs pool-owned, task-conditioned, uncertainty-aware utility estimates.

Synthetic costs include uncached input, cached input, output, and classifier/judge tokens. They are comparative assumptions, not provider prices. See the full report for formulas, rates, methodology, limitations, and reproduction assets.

Validation

• uv sync --locked
• 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
• uv run maturin develop --uv
• uv run pytest tests/ -v -m "not integration" — 116 passed, 2 deselected, 2 subtests passed
• make -C docs publish

All commands passed on Rust 1.96.1 in a Slurm compute job.