docs: microbench cross-tree index + synthesis-doc currency (org fixes)

[Lightheartdevs]

Addresses the repo-organization audit (Tiers 1–3). No data changes — navigation, taxonomy, and currency only.

Tier 3 — taxonomy (the headline org problem)

The 12-family agentic microbench (a model-behavior study) is split across benchmarks/ and hardware-tests/ by which GPU/quant a model happened to need, not by the question asked. Low-disruption fix (no directory moves, no broken links/history):

• NEW MICROBENCH-INDEX.md — one table of every 12-family microbench across both trees, plus disambiguation of the four "27B"s (AWQ / Q8 / FP8 / 35B-A3B sibling), and the one finding that holds across all of them (thinking net-negative).
• "Where this lives" taxonomy notes in the benchmarks/ microbench READMEs. (The 397B and 27B-FP8 entry notes ship in PRs Add MiniMax-M2.7-NVFP4 (N=5, TP=2): temp serving-trap + exhaustive-completer findings #33 / Qwen3.6-27B-FP8 full microbench N=5 — think vs no-think (clean FP8 redo) #34.)

Tier 2 — currency (synthesis docs were frozen at 2026-05-02)

• SCORECARD.md: added a 27B-quant scope banner (its "27B" = AWQ — kills the conflation risk), a newer-results summary (397B / Step / MiniMax / 27B-FP8), and marked the "FP8 higher-precision" follow-up partly-done.
• COMPARISON.md: "The FP8 re-run is the highest-priority follow-up" → done, cross-linked, with the confirmed result (thinking still net-negative at FP8).
• ROADMAP.md item 1: marked ✅ DONE for 27B; remaining FP8 work narrowed to Coder-Next + 35B-A3B.

Tier 1 — discoverability

• Indexed the two previously-unindexed hardware-tests entries (best-stack-followup, qwen3.5-397b-vs-step3.7-flash) in both the root README and hardware-tests/README.md tables.
• Added a MICROBENCH-INDEX.md row to the root "five-minute answers" table.

Merge notes

• References the 27B-FP8 (PR Qwen3.6-27B-FP8 full microbench N=5 — think vs no-think (clean FP8 redo) #34) and MiniMax (PR Add MiniMax-M2.7-NVFP4 (N=5, TP=2): temp serving-trap + exhaustive-completer findings #33) entries — merge those first so the links resolve.
• Light expected overlap with Qwen3.6-27B-FP8 full microbench N=5 — think vs no-think (clean FP8 redo) #34 on the README index tables (rows inserted at different anchors to minimize conflict). claims.yaml is not touched here (its merge-order is between Add MiniMax-M2.7-NVFP4 (N=5, TP=2): temp serving-trap + exhaustive-completer findings #33 and Qwen3.6-27B-FP8 full microbench N=5 — think vs no-think (clean FP8 redo) #34).

🤖 Generated with Claude Code