Add Test-Driven autoloop strategy

[Copilot]

Ships a second autoloop strategy alongside AlphaEvolve. Where AlphaEvolve drives optimization (evolve an artifact against a scalar fitness), Test-Driven drives specification (pin behaviour in a failing test, implement until green) — covering the API-porting and bug-fix workloads that don't fit AlphaEvolve.

Layout

Mirrors .autoloop/strategies/alphaevolve/ so the two feel like siblings. No workflow changes needed — strategy discovery already routes off program.md's ## Evolution Strategy pointer.

.autoloop/strategies/test-driven/
├── strategy.md
├── CUSTOMIZE.md
└── prompts/
    ├── write-test.md
    ├── make-green.md
    └── refactor.md

Files

• strategy.md — runtime playbook. 8-step loop: load state → pick target → red → green → refactor → evaluate → update Test Harness → fold through to default loop. Includes deterministic overrides (failing tests are obligations; 3× error → forced rethink-test), a 5-attempt cap that pauses the program with td-stuck: rather than thrashing, and the Test Harness schema (t{NNN} ids, status transitions, range-collapse compression at harness_size_cap).
• CUSTOMIZE.md — creator-time guide (parallels alphaevolve's). When-to-pick rules, copy commands, the ## Evolution Strategy pointer block to drop into program.md, marker-by-marker guidance, and a worked tsb-bugfix example.
• prompts/write-test.md — red-phase framing: domain-knowledge slot, framework-setup slot, validity checklist (must fail with a useful message, must not couple to implementation details), reasoning template.
• prompts/make-green.md — green-phase framing: minimum-change discipline, explicit don't-modify list, anti-patterns (overfitting, speculative generality, weakening tests to shrink the diff).
• prompts/refactor.md — refactor-phase framing, gated on a green suite. Refactor vocabulary, an explicit not-a-refactor list (renaming for taste, formatting, comment polish), skip conditions, and the hard rule that the test set must be identical before/after — if a "refactor" needs a test edit, it's a behaviour change and belongs in a red iteration.

All <CUSTOMIZE: …> markers are preserved as fill-in points; grep -R "<CUSTOMIZE" strategy/ should return nothing once an adopter resolves them for a specific program.