explore: resolve a solution by content_signature in solutions and curate

[cafzal]

The gap

certify's per_pick names the certified plan that beats a dominated finalist by content_signature, and its own next_steps says to go curate that point in its place. Neither explore solutions nor explore curate accepted a signature — both wanted solution_id.

So the prescribed move meant listing the whole exact overlay and scanning by eye. On shift_coverage_staffing that listing is 80,746 chars for 253 points: it overflows the response cap and spills to a file, putting the answer back on disk. That is the exact round trip per-pick verdicts were added to remove — the certificate produced the answer, and the surface made you go find it again.

Surfaced by driving the example end to end on current main: both curated picks came back dominated, and re-curating required the spill file to map signature → id.

Why this shape

curate already keyed remove and rename on content_signature. Pinning was the odd one out, so this is the action agreeing with itself rather than a new convention. feedback is the other precedent — it has taken "solution_id OR content_signature" all along.

Resolution is scoped to the run source/scenario selects: the same plan carries the same signature in the heuristic run and the exact overlay, so the caller says which one it means. An unresolved signature returns the surface's standard error and names the run it searched, because passing the wrong source is the likely mistake.

Layering follows the existing split: engine helper (explorer.resolve_signature) beside get_solution, routing prose in certify's next_steps, read-side teaching in solution_interpreter → Reading the Certificate, contract language in the tool docstring and the architecture tools table.

Tests

910 passed / 2 skipped. The one that matters is test_certify_routing_round_trips — it executes the certificate's own instructions verbatim (certify → curate the named signature from source="exact" → unpin the beaten one → re-certify) and asserts the second certificate reports every remaining finalist optimal. A per-pick verdict the user can't act on without reading files off disk is the defect being closed, so the test closes the loop rather than checking the parameter is accepted.