Replace the node score sentinel with an explicit valid flag

[rasros]

SearchNode.score was a float holding one of two sentinel values. Nothing read it as a quantity — every one of its 47 references was a < INVALID_SCORE validity check — but it read as a quality anyone could sort by, and three separate defects this week came from something sorting by it after it had stopped meaning anything:

• epoch_parents sorted its Pareto front by it, a no-op ordering
• select_survivors picked the elite by min(key=n.score), which would have raised at runtime
• node filenames led with it, so the best-named file on disk was not the artifact the evaluator chose

What it becomes

SearchNode.valid: bool — the candidate's markup parsed, it rasterized, and the measures computed. Not a quality judgement and not orderable: order=True is gone, because ordering by a constant is what caused the above. Candidates are ranked by dominance over their measures, which needs a population, so a single node has no position of its own.

Result.score becomes Result.measured: bool — whether the measures have been taken yet. That is all it encoded: a result crosses the queue once before anything is known about it, and Result.valid already existed alongside it.

ChainState.score is gone; it duplicated the node's.

What went with it

INVALID_SCORE, VALID_SCORE, valid_scores() and score_std(), plus the statistics derived from them — pool_score_std in stats.csv and the spread readout in the dashboard. A spread of a constant measures nothing. Crowding distance keeps an infinity sentinel, now math.inf, which is a genuine infinite distance rather than a disguised boolean.

Elitism is redefined. It reserved a slot for the best score; with no single best it now guarantees a member of the top Copeland tier survives, which is the same intent — the majority relation cycles and crowding distance sorts for spread, so a generation can otherwise evict everything unbeaten.

--resume-top and clean_runs.py read the older filename shapes, and plot_run.py's front plot now draws two objectives against each other rather than a metric against a score, since there is no score axis to draw.

Scale

28 files, 428 insertions, 544 deletions. The test migration was the bulk of it: roughly 200 call sites constructed nodes or results with a score. I first tried it as a bulk regex, which produced more failures than it fixed and mangled several lines outright; that was reverted and the files were migrated one at a time, running each before moving on.