e2e: the Chainsaw catch budget counts only the global legs, not the per-suite collectors #3754

hack/e2e-chainsaw/.chainsaw.yaml:57-59 records the worst-case diagnostics cost of a failed test as ~19m, and derives it from the two global error.catch legs: the previous-instance log read at timeout: 6m (:82) and the crust-gather snapshot at timeout: 13m (:133). The figure is exact for those two, and the block says outright that it exists so the arithmetic does not have to be rediscovered.

It does not count the catch collectors that individual suites add. Those run on the same failed test, in addition to the global legs. Counting them is easy to get wrong in the direction of undercounting, for two reasons that are worth stating before the table: a suite's catch can sit at test level (spec.catch) or on a single step (spec.steps[].catch), and a script op that omits timeout: does not run unbudgeted. It inherits timeouts.exec, which is 5m.

The enumeration below is (.spec.catch[]?, .spec.steps[]?.catch[]?) | select(has("script")) | .script.timeout over every hack/e2e-chainsaw/*/chainsaw-test.yaml, so it can be reproduced or refuted rather than taken:

So a failed harbor test can spend ~28m in diagnostics against the ~19m the block states, and a failed qdrant test ~26m.

The gap is not the 2m collectors. It is that harbor and bucket each carry one op with no timeout: at all, and an omitted budget is the largest single contribution while being the only entry a reader cannot see a number for. qdrant's second op is the same shape of surprise from the other direction: it is explicit at 5m, but it lives on a step rather than in spec.catch, so a count that walks only one of the two levels misses it entirely.

This matters because of what the same block says happens at the job ceiling: the E2E job is capped at 180 timeout-minutes, a cap ends the job cancelled, and the ~4m48s grace window is consumed by the earlier Collect report step, so both cozyreport.tgz and the job log are lost. That was measured, on runs 29557003429 and 29568562337. A cascade reddening many suites is the case the block already calls out as spending a large share of the job in diagnostics, and the per-suite collectors make each of those tests cost more than the recorded figure.

Three things worth deciding separately:

1. Whether the recorded number should include per-suite collectors, or whether the block should state that it excludes them and where to look instead. The block's own closing advice, "bound the number of snapshotted tests rather than shrinking these timeouts", points at the first.
2. Whether a catch script may omit timeout: here at all. Inheriting timeouts.exec: 5m is what makes harbor and bucket dominant while reading like the cheapest rows in the table.
3. Whether per-suite catch should be allowed at both nesting levels, given that the two have to be enumerated separately and a partial walk silently undercounts.

Pre-existing: harbor, qdrant, bucket and openbao all carry these at cfaf2e086. The two OIDC rows are the smallest entries in the table and are new; they change the total by 4m and nothing else about the shape of the problem. This is not a regression from a specific change. It is an accounting gap in a comment that is otherwise precise, and worth closing while the comment is still trusted.