cyber: bind the consequence verdict to the default path, un-stick the soften, surface difficulty

[larstalian]

Why

A rigorous review of the cyber worldgen / evolution / curriculum / dashboard turned up four problems that make the training signal untrustworthy or the pool illegible. The verifier is supposed to set the agent's ceiling, but it didn't actually bind the default training path, and the curriculum's "soften" mutation was dead code on the exact worlds it's meant to help. This PR fixes them before any GPU run, so the reward we'd be training against is honest, and makes the evolving pool readable in the dashboard.

This is the A + thin C slice from that review (honesty fixes + difficulty made visible). No new capability — it makes the existing one trustworthy.

What

Honest worlds (A)

• A benign request can no longer leak the flag. The default db handler used to serve any value by ?key, including a guarded (HIDDEN) one — so a sibling endpoint could make a world "winnable" without the intended exploit. The handler now filters guarded values, and verify.verdict sweeps every reachable benign endpoint and rejects the world if any of them leaks. (Closes the sibling-db key-lookup leak.)
• The consequence gate binds the default path. A task-less world now fails instead of passing blind; seed-gate rejections are logged; and the seed gate is wired into the notebook's training pool + held-out set (the evolve-time gate was already wired).
• The chain-collapse soften is reachable again. On networked worlds its relevance was always 0 (it keyed on in-process paths that never appear in requests_made). It now derives relevance from the agent's engagement with the public SSRF foothold — the one signal that survives the in-process proxy — floored above decoy-removal, so a chain-stuck agent actually eases the chain instead of cosmetically dropping a decoy. First loop-reachability test for it.

Legible pool (C)

• The pool now stamps each evolved world's solve-cost onto its lineage, and the dashboard reads + displays it. (It was reading an unrelated key that was always empty, so the panel never showed a difficulty.)

Testing

• 5 new regression tests:
  • benign db endpoint never serves the flag
  • whole-world verdict rejects a sibling leak
  • evolved snapshot persists world_difficulty
  • chain-soften is reachable from real agent reports
  • lineage node surfaces world_difficulty to the dashboard
• Full suite green: 984 passed, 17 skipped, 0 failed.

🤖 Generated with Claude Code