fix(seeds): enforce minItems in structured-output schema to prevent short-count failures

[jakepresent]

Problem

The seeds stage aborts when the model returns fewer items than requested per-cell batch, even by one. With larger sample_size values, this is statistically inevitable and the pipeline cannot proceed past seeds.

Repro (clean main, fresh artifacts dir):

# eval config based on examples.travel_planner_langgraph
seeds:
  prompt:
    model: { name: azure/gpt-5.4-mini, temperature: 0.7 }
    sample_size: 500  # 18-cell design grid → ~27-28 seeds per cell

Three consecutive runs failed at the same stage with the same shape:

ValueError: prompt generation returned 26 seeds for a batch of 27
ValueError: prompt generation returned 22 seeds for a batch of 23
ValueError: prompt generation returned 25 seeds for a batch of 26

Bumping max_tokens 5× (3000 → 16000) did not help, confirming this is not output truncation. The model was simply emitting one fewer item than the prompt asked for.

Root cause

seeds_response_schema() was generating a schema with no minItems constraint on the seeds array:

"seeds": {"type": "array", "maxItems": 2000, "items": item_schema}

The prompt asked for N seeds in natural language, but the schema told the model "0 to 2000 is fine." When the structured-output layer enforced the schema, returning N-1 was a valid response. gpt-5.4-mini did this consistently at batch sizes above ~10-15 per cell.

The strict equality check at seeds.py:668 then aborted the entire stage on the first short response:

if len(returned_seeds) < job.count:
    raise ValueError(...)

Fix

Thread min_items through seeds_response_schema() and pin it to job.count per call. The schema now enforces the count the prompt asked for, eliminating the under-generation at the structured-output layer instead of catching it after the fact.

This mirrors the existing pattern in _labels_response_schema(), which already pins both minItems and maxItems to count - so the precedent in this codebase is to use schema-level enforcement for exact-count structured outputs.

Validation

• Repro: confirmed broken on main. Three runs of sample_size=500 on examples.travel_planner_langgraph failed at seeds with the short-count error.
• Repro: confirmed fixed on this branch. Same config now generates the full 502 test cases (500 prompts + 2 scenarios) and the pipeline proceeds to rollout. Run is still in progress at the time of this PR; will paste final perf numbers in a follow-up comment.
• Unit tests: 4 new tests in SeedsResponseSchemaTest covering default behavior, min_items enforcement, and zero/negative-value handling. Full seed-related suite (105 tests) passes.

Why the strict check stays

I considered loosening the worker-layer check to accept-and-warn instead of raising on short count. Decided against it because:

1. The schema fix removes the failure mode at its source rather than papering over it downstream.
2. The strict check is still useful as a safety net - if the model violates the now-enforced minItems constraint somehow, we want to know rather than silently produce undersized batches.
3. Other stages (notably _labels_response_schema) already use the same pin-via-schema pattern, so this keeps the codebase consistent.

If there's appetite for a separate retry-on-short-count or accept-and-warn fallback at the worker layer, happy to do that in a follow-up.

[tangym]

With minItems now pinned to job.count but maxItems still at 2000, could the model over-generate? Probably unlikely, but curious.

[jakepresent]

Good catch @tangym - pushed a4169e4 that pins both bounds to job.count. Now the schema demands exactly N items per batch, matching the prompt's stated count.

Added two tests for the symmetric-bounds case and the non-positive override fallback. 30/30 seed-schema tests pass.