feat(content-refinement): reviewer decision bands + Devil's Advocate concession guard

[Ar9av]

What

Upgrades the content-refinement-agent simulated-reviewer loop with two deterministic mechanisms, both backward-compatible with the existing score_delta.py exit-code contract:

1. 0-100 decision bands with a principled target-met early halt.
2. A Devil's Advocate concession guard that stops the reviewer from sycophantically caving.

This is item #4 from the study of academic-research-skills, reimplemented to fit this repo (concepts only — no files copied; that repo is CC-BY-NC).

Why

The loop already had a 6-axis 0-100 rubric, anti-inflation caps, score deltas, and a prose-only DA protocol. Two gaps:

• The reviewer's decision was free-form and not tied to the numeric score — it could drift (e.g. "Accept" at overall 62), and the loop had no absolute quality target, only relative deltas.
• The DA concession rules ("concede only on strong rebuttal", "no consecutive concessions") were prose only — nothing stopped the simulated reviewer from quietly relaxing them and caving on a real weakness.

How

File	Role
scripts/decision_band.py	new — map overall_score → canonical band (Accept ≥80 / Minor 65–79 / Major 50–64 / Reject <50); importable + CLI
scripts/score_delta.py	extended — emits decision_band_prev/curr; adds target-met halt (exit 5)
scripts/concession_guard.py	new — deterministically enforces the DA concession protocol

Decision bands + target-met halt. Bands give the loop an absolute target on top of the relative accept/revert logic. When an accepted iteration reaches the Accept band (overall ≥ 80), score_delta.py returns the new exit code 5 (HALT_TARGET_MET) — promote the current draft and stop, rather than risk regressing it for marginal gains. Takes precedence over the plateau halt; --no-target-halt disables it (band still reported). Exit codes 0/1/2/4 are unchanged, so existing callers keep working.

DA concession guard. Reads a concession log (da_concessions.json) and re-derives which concessions are legitimate:

• Rule 1 — evidence: a concession is valid only at rebuttal_score ≥ 4. Caving at ≤3 is rejected; the finding is restored to standing.
• Rule 2 — no consecutive concessions (iron rule): a concession in a round immediately following another is rejected.
• A standing CRITICAL (unresolved and not validly conceded) → exit 1 (BLOCK) → host treats the iteration as REVERT regardless of rubric scores. A rejected non-blocking concession → exit 2 (WARN) → DA must restate. Exit 3 = schema error.

Testing

• py_compile clean on all three scripts.
• decision_band.py: 81→Accept, 74.6→Minor, 58→Major, 42→Reject; reads overall_score from a real score.json.
• score_delta.py: accept below band → exit 0 (Major→Minor); reaches band → HALT_TARGET_MET exit 5; --no-target-halt → exit 0; regression → exit 1. Regression-tested on the real workspace iter1→iter2 files: still ACCEPT_IMPROVED, exit 0, behavior preserved.
• concession_guard.py: clear→0/PROCEED, caving-on-critical→1/REVERT (concession rejected, critical stands), consecutive-concessions-on-majors→2/DA_RESTATE, unresolved-critical→1/REVERT, bad severity→3.

Docs

• references/reviewer-rubric.md — decision-band table + decision_band output field
• references/da-reviewer.md — concession-log schema + guard verdict→action table
• references/halt-rules.md — exit 5, decision bands, guard override of ACCEPT→REVERT
• SKILL.md — wired into Steps 1/2/5/6, Resources, and frontmatter description

Note: branched off main, independent of #8 (which touches literature-review-agent). No overlap.