feat: add safe exploratory training dataset builder

[shaggitza]

Summary

Adds a lightweight exploratory PR-to-training-JSONL workflow, deliberately isolated from publication-grade benchmark custody.

• hardens collect_prs.py as an offline-testable candidate-corpus collector with validated GitHub responses, strict UTF-8/JSON, controlled config errors, and no-clobber output;
• adds build_training_dataset.py to join a corpus with validated completed legacy reviews or v2 compatibility projections while skipping unlabeled/explicitly non-evaluable rows instead of inventing negatives;
• preserves legacy and v2 terminal status semantics, filters entrypoints by analyzer scope, and supports embedded, cached, or optionally fetched PR diffs;
• adds a private POSIX directory-FD publisher for race-safe exclusive output/cache publication without following symlink path components;
• validates full Git OIDs and confines deterministic cache names to the configured diff directory;
• documents the separation between exploratory generation, frozen analyzer comparison, and publication-grade truth.

Safety and scope

• No frozen corpus, review, adjudication, production, or v2 custody artifact is modified.
• Output and fetched caches are exclusive-create only; there is no overwrite/--force mode.
• Frozen artifacts, publication roots, input aliases, existing destinations, symlinked path components, and unsafe cache aliases are rejected.
• JSON is strict: duplicate members, non-finite numbers, invalid UTF-8, and non-serializable output fail closed.
• Completed labels require their expected fields and evidence; malformed statuses or duplicate identities fail instead of becoming empty targets.
• The secure publication helper is private to the exploratory tools and intentionally relies on Linux/POSIX filesystem semantics.

Validation

• 40 focused collector/builder tests passed, including path races, traversal, cache aliases, label/status validation, strict JSON/UTF-8, malformed GitHub responses, and first-fetch/cache repeatability.
• All benchmark tests passed locally excluding only tests/benchmarks/test_pilot_typed_run_v3.py; four expected installed-runtime tests skipped.
• Ruff format/check passed for all changed Python files.
• Strict mypy passed for all changed Python files.
• Two frozen-corpus smoke builds were byte-identical: 58 examples, 2 skipped labels, 58 missing diffs; SHA-256 d09e43404c71b8d769eeb94f3e693f2ec4426d3935839b477ec5e8ceafc06656.
• Seven protected frozen artifacts remain byte-identical to HEAD.
• Terminal independent acceptance review returned PASS / merge-ready.

Local limitation

The platform-specific native pilot broker suite remains unavailable locally because this host does not implement os.posix_spawn(..., setsid=True). GitHub Actions is the authoritative run for that existing suite.