feat: wire diagnostic costs into profile triage pairs

[acidkill]

Completes the in-progress wiring of caller-verified diagnostic costs (VCR405, #307) into the paired-trial harness. Builds on an uncommitted work-in-progress edit that previously crashed ProfileTriageSpec construction and failed the local suite (8 errors). Also fixes a pre-existing argv-order divergence so rank-plus-observation triage commands are intercepted by the bridge instead of silently falling back to the real interpreter.

Change

• Case profiles may declare optional triage.diagnostic_costs for supplied hypothesis IDs only (plain-string enum tokens, loader-validated).
• Both arms receive the same cost disclosure; the treatment arm is told costs may guide next-check ordering only, are never causal-likelihood evidence, and never waive required checks.
• ProfileTriageSpec gains a validated diagnostic_costs field (plain strings only; they are embedded verbatim in the generated shim).
• command_for, the intercepted python shim, the bridge request contract, and _decide forward the costs end to end, so cost-bearing triage requests stay bridge-counted instead of silently falling back to the real CLI, and the tokens reach the production triage_failure decision state.
• _triage_argv now places observation flags before --rank-hypotheses, matching command_for and the shim; previously the exact command for a rank-plus-observation profile was rejected by the shim's parser and fell back to the real interpreter, bypassing bridge call counting and receipts.
• Profiles without costs keep identical prompts, commands, requests, and decision state.

Evidence (local, offline)

• 10 new tests in tests/test_pilot_profile_diagnostic_costs.py: loader accept/reject, equal-arm disclosure, supervisor argv construction, spec validation, bridge request matching/forwarding, shim interception without fallback, and rank-plus-observation command order (verified to fail against the pre-fix ordering).
• Updated existing doubles/request bodies for the new wire-contract key.
• Full suite: 858 tests OK (43.1 s); git diff --check clean.
• Ledger entries: VCR408 in TASKS.md and PILOT.md.

Local results only — the required hosted CI check must be green before merge. No trial was run; no efficacy, delivery, or release claim. No historical trial is rescored or rerun.