eval: cover Agent Spec launch method selection

[schickling-assistant]

Problem

The proposed Agent Spec start/resume lifecycle has no model-free behavior cell covering method selection, exact-session continuation, unavailable-method policy, and compatibility with the existing single argv form.

Goal

Add contract-first expected-red coverage for the proposed syntax while keeping the current portable runner and canonical AGENT-SPEC.md unchanged.

Decisions

• Cover five closed-set cases: explicit start, exact resume, unavailable refusal, declared start fallback, and legacy argv.
• Treat start/resume availability as declaration-level behavior; the probe never inspects provider transcript stores.
• Require unavailable-without-fallback to emit the exact resolved-identity diagnostic, exit nonzero, launch nothing, and leave session records unchanged.
• Keep the four proposal cases at RED/P01 against the published runner. Expected reds are classified evidence, not conformance passes.
• Preserve the legacy case as a positive control.
• Mark the syntax as proposed and unratified; do not update AGENT-SPEC.md.

Verification

• Published runner 0fed14b: four proposal cases RED/P01, legacy case PASS, PRODUCT-RED, zero residue; 5/6 judges pass and the run step fails as intended.
• Held-out st2 candidate add61394c335092eaed6d85e00d7befe8b05c023: all five behavior cases pass, PRODUCT-GREEN, zero residue. The outer eval fails only the deliberately frozen current-runner classification map.
• bash cells/launch-method-selection/judges/self-test.sh: 6/6 contract mutations rejected.
• 229 shell scripts parse.
• Preflight safety, model policy and mutations, run history, retired surfaces, event-first, harness contract, closed-set mutations, overnight policy, PII/history, catalog freshness, and git diff --check pass.
• bin/check-corpus.sh remains red on the unchanged weird-git-setup fixture: its intentional alternate commit identity is rejected by the globally installed assistant-identity commit hook. The fixture and check are unchanged from main, and the failure predated this change. The hook was not bypassed.

Complexity

Medium. The cell is model-free, but it exercises five isolated catalogs and adversarial synthetic session records with explicit cleanup.

Concerns

• The field names and defaults are not ratified; the cell tracks the proposal rather than defining the canonical spec.
• The current CLI and receipts do not expose parser-level preservation of resume.session; upstream model tests own that assertion.
• Open eval PRs test: freeze Agent Spec field-change contract matrix #67 and eval: prove stable agent identity and presentation metadata #70 also modify generated catalog/evidence files and may require mechanical conflict resolution.

Friction & bottlenecks

The repository did not ignore the tmp/ directory used by the required Axe worklog. This PR adds the missing ignore rule so local tooling metadata cannot enter the publishable PII surface.

Follow-ups

When a ratified implementation becomes the published eval runner, update the frozen case manifest from the four P01 expected reds to the accepted behavior map.

References

Refs compoundingtech/st2#124.