docs: free-ceiling research report (deep-research, cited)

[ssamba1]

Adds docs/free-ceiling-report.md — a cited synthesis from the project's deep-research harness on the realistic ceiling of a zero-out-of-pocket closed-loop humanizer.

Method: 5 search angles → 22 sources → 104 claims → top 25 through 3-vote adversarial verification (9 confirmed, 16 killed).

Findings:

• Published ceiling (92–97.6% ASR) requires GPU fine-tuning (MASH / StealthRL / AuthorMist); no verified training-free method reaches it.
• Ceiling per detector tier: high vs local open detectors (in-loop; the training-free closed loop untell mirrors reports −88% TPR@1%FPR), reachable-but-fragile vs ZeroGPT, unprovable vs GPTZero / Originality / Turnitin (can't put them in the loop for free; Turnitin has no public API; GPTZero ships a dedicated anti-humanizer classifier).
• untell's actual inference-only regime is unmeasured in the literature — the Implement humanize v1: detector-feedback skill + eval harness #1 free thing worth measuring next.
• Ranked free moves + an honest "what free cannot do" + killed folklore (ZeroGPT-is-useless, Binoculars-collapses-on-GPT-4, vulnerabilities-don't-transfer — all refuted in verification).

Honest-by-design, consistent with the rest of the repo.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.