workflow-fix #777: flip dual-DV continuous-companion polarity (fixed +/- margin PREFERRED, logp_pos_mean OPT-IN)

[superkaiba]

Closes task #777.

Summary

Corrects the dual-DV continuous-companion guidance in the project's
measurement-validity rules. The PREFERRED non-saturating continuous DV
is now the teacher-forced fixed positive-vs-negative completion margin
(fixed answer set across all contexts → no selection-on-outcome bias);
the length-normalized log-P of judged-positive on-policy completions
(logp_pos_mean) is flagged as an OPT-IN alternative that carries
selection-on-outcome bias and FAILED task #722's ρ(DV, rate) > 0
validation gate for 3 of 4 behaviors (ρ ≈ −0.3).

Evidence

Task #722's eval_results/issue_722/tf_margin/margin_chain.json:

• margin ρ(DV, rate): refusal +0.56, sycophancy +0.40, broad_em +0.31 — all positive.
• logp_pos_mean ρ(DV, rate) ≈ −0.3 for 3 of 4 behaviors — FAILED the same gate.

Diff

10 markdown files (0 .py, 0 configs), +65/-13 lines total across 2 commits:

• d30d1fd600 — polarity flip in CLAUDE.md + .claude/rules/{marker-leakage-measurement,llm-judging}.md + 5 enforcer agent files (planner.md §6, critic.md Stats lens, analyzer.md step 3, interpretation-critic.md Lens 1, clean-result-critic.md dual-DV lens).
• 883352fcc9 — folded-in Codex Minor CONCERN: bump 18→20 guideline count references to acknowledge the new § E2 continuous-companion-DV recipe alongside the 18 judging guidelines.

5 MUST-NOT-CHANGE invariants preserved

1. On-policy judged RATE stays PRIMARY.
2. ρ(DV, rate) > 0 validation gate intact.
3. Task #558: persona-panel read on the #543 Phase-2 adapters #432→issue-621: rank-1 LoRA read/write vector implants #456 teacher-forcing caveat intact.
4. Graded-judge primary-for-ranking-targets extension intact.
5. Marker three-space recipe (Log and analyze ALL THREE spaces) untouched.

Test plan

• scripts/select_step9c_tests.py --base main touched-file scope: 2318 pass / 5 fail / 6 skip.
• All 5 failures verified pre-existing on main (byte-identical evidence): 3× the #745 scripts/issue744_dump_and_stream.py:87 bare-dotenv lint (tracked separately), 2× HF-Hub network-state audit-claim tests.
• 4 polarity-survivor greps empty in all 10 touched files (no old-wording survivors).
• 4 invariant-preservation greps present.
• Both code-reviewers PASS-class (Claude PASS, Codex CONCERNS→Minor 18→20 count folded in).

🤖 Generated with Claude Code

[superkaiba]

Content already on main via concurrent-session rebase-merge (see main commits 588370960f polarity flip + 02e8e01cd7 count fix). This PR's original commits d30d1fd600 + 883352fcc9 were rebase-rewritten during that merge; gh pr merge --rebase now returns "branch can't be rebased" because the commit content is semantically present on main. Closing rather than deleting the branch so the artifact-confirmed history is intact.