D086 - Rank blockers by shaders blocked, within an effort tier

decided · 2026-08-19, revised 2026-08-20

The single judgement in coverage.rs, and it decides what gets worked on.

An instruction appearing ten thousand times inside one shader blocks exactly one shader. An instruction appearing once each across four hundred shaders blocks four hundred. Ranking by raw frequency puts the first at the top of the list and buys a week of work for one shader.

Occurrence count is kept, as context and as a tiebreak, but it is not the ranking.

The same entry separates undecodable from decodable-but-untranslatable, because they are different work: the first is usually a fix to the table above, the second is a feature in a translator that does not exist yet. Collapsing them sends someone to the wrong file.

Revised: effort is the first key, shaders blocked the second

Confirmed with input, and extended, the first time the list had real data in it.

Ranking by shaders blocked answers what would help most and says nothing about what is reachable. The two came apart immediately: the top entry was an export, blocking two shaders and needing an entire render-target model, while the bottom was an ordinary multiply-add that took twenty minutes. A single ordered list puts a week of work above a morning's and offers no way to tell them apart - and following it would have produced nothing that day.

So blockers now sort by effort first: ordinary work above anything waiting on a subsystem, and shaders blocked within each tier. The output separates the tiers too, because a reader scanning for what do I do next should not have to know that the list silently changes meaning partway down.

Two tiers, not a score. Blocked-divided-by-effort would rank them precisely and the precision would be invented - nothing here can measure effort, and a ratio built from a guess reads like a measurement. Two tiers claim only what is known.

The tier is not a new hand-maintained table. It comes from the translator's existing blocked-instruction list, which already records why each refusal stands and already has a reason to stay accurate. The shader crate can see what blocks a shader and not what it would cost; the translator knows the reason; the caller joins them. Nothing new to keep current, and no second place for the two to disagree.