docs: validity boundaries, ROADMAP, FIELD-REPORTS, quant-contribution path

[Lightheartdevs]

Summary

Phase 4 of the docs reorganization. Translates the substantive criticism from the recent LocalLLaMA discussion thread into validity-boundary docs and a contribution-pathway scaffold. No changes to data or methodology — purely about making the study's scope explicit and giving external contributors a clear way to fill the gaps it doesn't cover.

• README.md + COMPARISON.md — hoisted "operating point" callout (4-bit AWQ Cyankiwi on 2× RTX PRO 6000 Blackwell; other quants/VRAM/hardware not characterized) to entry-point visibility. Stops the "what quant?" and "what about my hardware?" questions before they're asked.
• COMPARISON.md — new "What this benchmark doesn't characterize" section before the Drilling-deeper table. Five paragraphs covering: other quants of the same models, other VRAM tiers, other hardware classes (Mac M-series), other languages (Python-only Phase 1), single-rig hardware variance.
• KNOWN-LIMITATIONS.md — expanded "Quantization specificity" subsection with a new "Cyankiwi 4-bit AWQ field reports" block. Acknowledges multiple practitioner reports that these specific quants underperform official FP8 / Unsloth UD4 of the same base models. Defends the within-quant comparison without defending absolute capability claims at higher precisions.
• microbench-phase-b/findings.md — extended Recommended follow-ups from 3 → 6 items; added FP8 re-run, M-series Mac sibling study, and language-mix expansion.
• New `ROADMAP.md` — consolidated open questions and contribution opportunities from across all findings docs. 10 prioritized active follow-ups with [contributor-welcome] flags on the 4 items external contributors can take end-to-end.
• tooling/ADDING-A-MODEL.md — added "Two contribution shapes" section clarifying that "same model, different quant" is a valid contribution path, not just adding a wholly new model.
• New `FIELD-REPORTS.md` — template for collecting voluntary practitioner reports of model behavior on real workflows. Complements the structured benchmark data with anecdotal-but-specific evidence.

What this PR is NOT: a response to specific commenters, a change to the benchmark data, or a folder restructure. The discourse is ephemeral; the docs are permanent. This translates the signal into validity-boundary docs and roadmap items.

Test plan

• Render check on the new "What this benchmark doesn't characterize" section in COMPARISON.md — verify the in-page anchor link from the operating-point callout resolves correctly
• Render check on ROADMAP.md and FIELD-REPORTS.md
• Verify the cross-references between COMPARISON.md, KNOWN-LIMITATIONS.md, ROADMAP.md, FIELD-REPORTS.md, and ADDING-A-MODEL.md all resolve (CI link-check should report this)
• Confirm `check-links` workflow passes with `checked=867 missing=0`
• 5-minute test (informal): a reader arriving via a public link should hit the operating-point framing within 30 seconds and understand what isn't measured within 90 seconds

🤖 Generated with Claude Code