fix(kb): stop storing what_worked/what_failed rows as empty strings

[rpoornac]

What breaks

what_worked / what_failed rows reach the KB with every value replaced by "".

The Coordinator is the only producer of these rows and stamps them with {name, extra_server_args, extra_envs, gain_pct, source} (writeback._build_recipe_payload). The store then normalised each row down to the arbor pair — _normalise_str_dicts(what_worked, ("description", "measured_impact")) — a vocabulary the Coordinator never writes. Nothing translated between the two, so every value was dropped:

sent:   {"name": "mtp_on", "extra_server_args": "--speculative-num-steps 3",
         "extra_envs": {...}, "gain_pct": 4.2, "source": "https://pr/123"}
stored: {"description": "", "measured_impact": ""}

This is not a missing field. The row count survived, so the KB looked populated while carrying nothing.

What that cost

prelude._warm_recipe_proven_items reads name and skips rows without one, so it always returned []. The already_proven list handed to the scout in phases/internal therefore carried only the current session's accepted variants. Anything proven in an earlier session was invisible, and the scout re-researched it — the KB's central promise, not paying out.

The fix

Finding and Failure now carry the producer's fields as hyperloom superset fields — the same convention as the existing Pitfall.severity, which arbor consumers ignore — and a new _normalise_experience preserves them instead of discarding them.

Two compatibility properties hold by construction:

• description falls back to name, so an arbor consumer gets operator-readable text where it previously got "".
• Superset fields serialise only when populated, so a row carrying just the arbor pair is byte-identical to before. The wire shape is unchanged for rows that never had these fields.

Why no test caught it

Both halves of the seam were already covered: one test asserts writeback stamps name, another asserts prelude reads it. But the prelude test hand-injected the row, so nothing exercised the persistence step between them — exactly where the data died.

The two added tests write through LocalRecipeStore and assert on what comes back out. Against main they fail with the bug in plain view:

assert worked[0].get("name") == "mtp_on"
E   AssertionError: assert None == 'mtp_on'
E    +  where {'description': '', 'measured_impact': ''}.get

assert {p["name"] for p in proven} == {"mtp_on", "fp8_kv"}
E   AssertionError: assert set() == {'fp8_kv', 'mtp_on'}

Scope

The explore dedup pre-fill stays inert for now: it fingerprints from extra_server_args / extra_envs, which the Coordinator does not record on what_failed rows. Storage no longer stands in the way; that remaining fix is producer-side.

Verification

• 569 passed across the recipe_kb / sedimentation / warm-start suites.
• Full suite: 22696 passed. The 9 failures are pre-existing on main at the same base (2cc382a83) in test_agentx_repair, test_external_multi_node, test_preflight_auth_override and test_supervisor — none touched here. The one that differed between worktrees (test_dotenv_fallback_ignores_arbitrary_cwd_dotenv) was traced to a stray .env in the local worktree: it fails there with main's code too.
• ruff check src/hyperloom clean.