Second looks are detected by observation overlap, not prevented

Every other guard keys on identity: the spec hash catches a re-run of the same Strategy. A mechanism Hypothesis has no spec, so "does regime predict fade success?" and "is fade success conditional on trend state?" are the same question in different words, and nothing else would notice.

Detection reuses the bookkeeping ADR-0011 already requires. A new Hypothesis whose consumed observations overlap a resolved one on the same axis beyond a declared threshold is refused until it carries an explicit link — either supersedes, stating what changed, or a stated distinction explaining why it is a different question. The threshold is configuration.

Considered options

• Nothing; rely on alpha cost. Registration already costs, so re-asking is self-limiting. Rejected because rewording is cheap relative to a budget, and the donor programme's own near-duplicates (H4-ORDERFLOW -> H4-ORDERFLOW-v2) were linked because a person chose to, not because anything required it.
• Hard block on overlap. Strongest guarantee, and it would refuse legitimate replications, corrections after an engine fix, and genuinely new mechanisms examined on the same data. A guard that costs more than it saves.

Consequences

This is deliberately a soft gate. No rule can distinguish a genuine follow-up from a reworded retry, and one that claimed to would produce confident wrong refusals. What it does is make a second look visible and deliberate: an operator may still write "this is different because...", but they must write it, in the register, beside the thing it overlaps. That is a materially different act from not noticing.