Evidence-based BLAST runtime and cost estimate

Motivation

The submit page had a cost-estimate stub but no defensible runtime model. Static program multipliers would produce precise-looking numbers without evidence, especially across databases, query sizes, VM families, and node counts that scale differently.

User-facing change

The Runtime summary now requests an estimate only when inline FASTA, database letter count, and a non-zero workload pool are known. It displays the robust median duration, interquartile range, comparable sample count, confidence, and approximate compute cost. Input changes are debounced by 750 ms and estimates are cached for 60 seconds.

While changed input is still inside the debounce window, the previous input's estimate is hidden and the summary reads Calculating. An estimate for one query/cluster combination is never shown beside a newer form state.

Fewer than three comparable completed jobs returns an explicit insufficient_samples state. No duration or cost is shown in that case. The estimate is informational: it is not part of validation, preflight, submit payloads, admission, budgets, or Service Bus messages.

API change

POST /api/blast/runtime-estimate accepts a typed read-only request. The service scans at most 500 existing rows in the selected scope and keeps only terminal-success samples with matching program, normalized database name, and workload VM SKU plus complete query/database size, node-count, and compute duration evidence. Extreme instrumentation/unit errors beyond a fourfold median band are excluded. Repository failure degrades to HTTP 200 with available=false.

No historical row, cost preference, submit state, queue, or completion event is written.

The pre-existing /api/blast/cost-estimate lab-tool stub is unchanged.

Validation

• uv run pytest -q api/tests/test_blast_runtime_estimate.py - 5 passed.
• Runtime estimate plus existing cost route/estimator suites - 18 passed.
• Frontend runtime model/display tests - 5 passed.
• npm --prefix web run build - passed.
• Protected Service Bus source hashes remained identical to the pre-expansion baseline.