* docs: PR 1.1 — relational leakage audit on alpha bundles

Reproduces the ChatGPT v2 finding on the actual 5000-lead alpha bundles
(`release/{intro,intermediate,advanced}/`). All three public tiers
reconstruct `converted_within_90_days` at 100% via every probed path;
the join-derived LR / HistGBM model also achieves AUC = 1.000 in 5-fold CV.

Adds:

- `scripts/probe_relational_leakage.py` — five deterministic reconstruction
  paths (A direct, B opportunity won, C customer exists, D subscription
  exists, E B∨C∨D) plus a bonus public-relational-only model. The
  deterministic reconstruction is exposed as a pure function
  `deterministic_relational_reconstruction(...)` designed to lift verbatim
  into `leadforge/validation/leakage_probes.py` in PR 3.1.
- `tests/scripts/test_probe_relational_leakage.py` — unit test on a
  hand-built four-lead toy bundle covering each path independently, plus
  a path-A-collapses-when-label-dropped test that pre-validates Phase 2's
  fix shape.
- `docs/release/v1_current_state_audit.md` — per-tier evidence table,
  explicit G4.1-G4.6 verdict (all FAIL), pointer to PR 2.1 as the
  structural fix.

Also reaffirms G1.1 (dataset name `leadforge-lead-scoring-v1`, already
locked via PR #61's milestone rename) in `v1_acceptance_gates.md`, and
checks off Phase 1 in `.agent-plan.md`.

This PR is evidence + docs only. The structural fix lives in PR 2.1.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* review: address self-review findings on PR 1.1

Brutal self-review identified 15 issues; this commit addresses the
correctness/credibility ones. Summary:

Script (scripts/probe_relational_leakage.py):

- Assert lead_id uniqueness in deterministic_relational_reconstruction
  (a validator must refuse non-unique keys, not silently misalign).
- Make customers/subscriptions optional inputs — empty frames now
  accepted, so the same script runs cleanly on Phase-2-fixed bundles.
- Branch the agg dict properly — was using has_close_outcome flag
  *inside* a lambda that pandas would have crashed before invoking.
- Add Phase-2-success ablation: re-run deterministic probes on a
  redacted-in-process view to surface the post-fix shape *now*.
- Bonus model gains a without_close_outcome_aggregates variant —
  the with-aggregates variant is trivially perfect because
  any_closed_won is Path B aggregated; the without variant is the
  load-bearing probe. Refactored to sklearn Pipeline (removes
  string-based scaler dispatch).
- Add G4.4 snapshot-window probe directly on event tables. Empirical
  result: alpha bundles PASS G4.4 (90-day horizon already bounds
  timestamps). The audit doc now corrects the prior "FAIL (assumed)"
  framing.
- Add --max-accuracy flag for Phase-2 CI gating (exit 2 on threshold
  violation).
- Rewrite docstring to clearly distinguish four orthogonal evidence
  channels.

Tests (tests/scripts/test_probe_relational_leakage.py):

- Add _binary_metrics coverage (perfect / all-negative / mixed) — the
  function had divide-by-zero handling and F1 logic with zero tests.
- Add lead_id-uniqueness rejection test.
- Add CLI smoke test on release/intermediate (skip if not present).
- Add --max-accuracy gate failure test.
- Drop the synthetic edge-case lead — toy bundle is now 3 leads
  cleanly covering A/C without B/D.

Audit doc (docs/release/v1_current_state_audit.md):

- Distinguish Path A (label in cleartext) from Paths B-E (joins).
- Add Phase-2 ablation table showing collapse to (1 - conv_rate).
- Report bonus model in two variants; AUC = 1.000 in BOTH variants
  is a substantive new finding (non-trivial relational features
  carry the leak independently of close_outcome — i.e., omitting
  customers/subscriptions tables is structurally required).
- Correct G4.4 verdict to PASS (with caveat about within-horizon
  content still leaking).
- Note instructor companion behavior is expected.
- Stop reporting AUC on deterministic 0/1 paths.
- Remove "sanity floor" framing for Path A (understated the violation).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* review: address Copilot inline comments on PR 1.1

Five real issues from the automated review:

1. Path A (label cleartext read) used .astype(bool), which maps NaN to
   True. Routed through pandas' nullable "boolean" dtype with explicit
   fillna(False) so partially-redacted bundles don't report phantom
   positives. Same fix applied to y_true construction in probe_bundle.

2. _binary_metrics F1 returned NaN when precision OR recall was exactly
   0.0 — should be 0.0 (zero-skill is well-defined; NaN should signal
   undefined). New logic: NaN only when precision/recall is itself NaN;
   0.0 when both are zero; harmonic mean otherwise.

3. Module docstring promised exit 1 on "shape errors" but main() only
   caught FileNotFoundError. Broadened catch to (FileNotFoundError,
   KeyError, ValueError); error message now includes exception type.

4. probe_bundle() previously bailed with an "error" key when
   converted_within_90_days was absent. Post-Phase-2 bundles will be
   exactly that shape. Now reports path_prediction_rates (fraction of
   leads each path flags positive) unconditionally — useful as a
   sanity view both before and after the fix. The early-return is now
   a "note" key rather than "error", with a message pointing the
   reader at the prediction-rates view.

5. New tests:
   - F1 zero-precision-and-recall yields zero, not NaN
   - Path A with NaN-bearing label column maps NaN -> False

Total: 10 tests, all passing. Lint/format/mypy clean.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>