fix: return CollectionStats from collect_results, unify zero-output check (#95)

[Muizzkolapo]

Summary

• First action failing all items (e.g., invalid API key → 401) was incorrectly marked completed_with_failures (PARTIAL) instead of failed (ERROR) because the initial pipeline had no zero-output check
• Root cause: collect_results() computed success/failure counts internally but discarded them — callers had to re-derive the information independently, and the initial pipeline didn't
• Fix: collect_results() now returns (output, CollectionStats) — both pipelines use the same stats to make the same decision

Changes

• result_collector.py: Added CollectionStats dataclass, changed return type to tuple[list[dict], CollectionStats]
• pipeline.py: Uses stats.failed instead of re-scanning results list for FAILED items
• initial_pipeline.py: Adds zero-output check using stats (was completely missing — this was the bug)
• Tests: Updated ~20 call sites to unpack new return type

Root cause analysis

Two code paths both call collect_results():

1. StandardStrategy → pipeline.py — had a zero-output check (PR feat: pre-flight validation and dependency-aware execution guards #67)
2. InitialStrategy → initial_pipeline.py — had no check at all

When the first action (initial strategy) failed all items with 401, it wrote empty output and returned successfully. The executor called _handle_run_success → _resolve_completion_status found failed dispositions → set completed_with_failures. The circuit breaker didn't fire because completed_with_failures is not a failure status.

Mitigates #95 — downstream actions no longer execute after the first action fails entirely.

Test plan

• ruff check . — clean
• ruff format --check — clean
• pytest — 4300 passed, 2 skipped, 0 failures
• Manual: first action with invalid API key → status failed → circuit breaker skips all downstream → tally shows 0 OK | 0 PARTIAL | N SKIP | 1 ERROR