fix: harden worker prompt against context compression

[jflowers]

Summary

Hardens worker.md agent prompt against LLM context compression by restructuring critical behavioral constraints. Fixes #49.

The worker's file reservation enforcement, progress reporting requirements, and learning storage constraints were in a separate "Constraints" section vulnerable to being summarized away by context compressors. This restructuring integrates each constraint inline with the checklist step it governs, making them atomic units that compressors must keep or drop together.

Key changes:

• Constraints co-located with their checklist steps using MUST/NEVER language
• Reservation failure recovery path added (STOP + comms_send to coordinator)
• Separate ## Constraints section eliminated
• File reduced from 27 to 20 lines

How to Test

# Run the content-verification test
go test ./internal/agentkit/ -run TestWorkerPrompt_HardenedStructure -v

# Run the full suite
make check

Verify structurally:

• worker.md has no ## Constraints heading
• Step 3 (comms_reserve) includes "NEVER edit unreserved files" and reservation failure recovery
• Steps 5 and 6 use "MUST" language for progress reporting and learning storage
• File is 20 lines (under 35-line target from design decision D4)

How to Demo

cat internal/agentkit/content/agents/worker.md

Compare the before (separate Constraints section with flat bullets) to the after (all constraints inline with numbered steps). Each step now reads as "do X — MUST/NEVER Y" rather than having the constraint in a disconnected section.

Key Files Changed

File	Change
internal/agentkit/content/agents/worker.md	Hardened prompt: constraints inline, MUST/NEVER language, reservation failure recovery, Constraints section removed (27→20 lines)
internal/agentkit/agentkit_test.go	New TestWorkerPrompt_HardenedStructure — structural verification test for the hardened prompt
openspec/changes/harden-worker-prompt/	OpenSpec change artifacts: proposal, design, delta spec, tasks

This PR was generated by /uf.finale (AI-assisted).