write results.json for structured agent consumption, harden crash diagnostics

[shichangs]

Summary

Fixes #64

• train.py: writes results.json after evaluation with the same metrics already printed to stdout. Gives agents a structured, parseable results channel instead of relying on grepping free-form stdout from run.log.
• program.md: instructs agent to read results.json first (fallback to grep). Crash diagnostics now use filtered grep -i "error|exception|traceback" instead of raw tail -n 50, reducing the surface for indirect prompt injection via training output.
• .gitignore: adds results.json and run.log as runtime artifacts.

Existing stdout output is unchanged. No new dependencies (json is stdlib).

Similar to #79 but rebased on current main with no merge conflicts.

Test plan

• Run uv run train.py and verify results.json is written with correct metrics
• Verify stdout output is unchanged
• Verify results.json values match stdout summary
• Simulate a crash (e.g. syntax error in train.py) and confirm results.json is NOT written

🤖 Generated with Claude Code

[shichangs]

cc @johnwaldo (issue author) @mvanhorn (PR #79 author) — would appreciate your review. This takes a similar approach to #79 but rebased cleanly on current main.

Offline verification

The results.json output can be validated without a GPU:

import json

# Simulate what train.py writes
results = {
    "val_bpb": 0.9979,
    "training_seconds": 300.1,
    "total_seconds": 325.9,
    "peak_vram_mb": 45060.2,
    "mfu_percent": 39.8,
    "total_tokens_M": 499.6,
    "num_steps": 953,
    "num_params_M": 50.3,
    "depth": 8,
}
with open("results.json", "w") as f:
    json.dump(results, f, indent=2)

# Verify roundtrip
with open("results.json") as f:
    loaded = json.load(f)
assert loaded == results
assert isinstance(loaded["val_bpb"], float)
assert isinstance(loaded["num_steps"], int)
print("✓ results.json roundtrip OK")

The key behavioral changes:

1. Success path: agent reads results.json (structured JSON) instead of grepping stdout → no injection surface
2. Crash path: agent uses grep -i "error|exception|traceback" to filter relevant lines before falling back to tail → reduced injection surface