docs: debloat prose and normalize ADR field formatting

[qflen]

Debloat pass across the docs: break prose walls into scannable bullets, trim connective tissue, and fix a formatting bug in DECISIONS.md. All technical facts, file paths, ADR refs, metric names, and numbers preserved.

Per file

• DECISIONS.md — fixed a render bug: 19 of 31 ADRs (014-031) had **Status**/**Context**/**Decision**/**Consequences** on consecutive lines with no separator, so GitHub collapsed them into one run-on line. Normalized all 31 ADRs to blank-line-separated fields. Also converted the crammed 7-item problem lists in ADR-020/021 to bullets. Content verified byte-identical apart from those two edits.
• ARCHITECTURE.md — every dense prose block (Storage, Ingestion, Metrics, Cluster, the ~25-line Downsampler paragraph) rewritten as bullets; em-dashes removed and prose tightened for review. 361 → 326 lines, ~13% fewer words.
• CHANGELOG.md — 10-13-line run-on bullets tightened; bold lead-ins for scannability.
• PERFORMANCE.md — the "characteristics" section de-walled; anti-entropy/rebalancing sub-bulleted. Tables untouched.
• PROTOCOL.md — resolution-selection and backfill walls tightened; endpoint specs untouched.
• README.md — already clean, no change.

Formatting only. No behavioral or API content changed.