docs: record data-layer D5 concurrency progress

[StackOverFlow11]

Summary

Docs-only progress sync recording Data Layer D5 (opt-in bounded concurrency + one shared global rate limiter) as implemented. No code change.

PR #51 merge status

MERGED into main (mergeCommit d4b665d). D5 code accepted by Codex; this PR only updates documentation to match.

Changed files (3, docs only)

• docs/data/data_layer_contracts.md — D5 recorded as implemented (opt-in bounded concurrency for updater/cache-warm fetches guarded by one shared GlobalRateLimiter; default max_workers=1 keeps serial; dense per-symbol daily endpoints can use workers while snapshot / index_weight paging / intraday stay serial; cache=raw endpoint SoT / PanelStore per-run artifact boundary unchanged); section-4 title gains / D5; the unimplemented block moves to D6+ (endpoint schema registry still deferred; D6 = PanelStore append/partition / optional materialized derived panel store, only if factor research needs reusable derived panels). Only bounded global-rate-limited concurrency was completed — no other "D5+" idea is claimed.
• CLAUDE.md — PR feat(data): D5 opt-in bounded concurrency with a global rate limiter #51 added to the merged-PR list; D5 progress bullet; quality gate 656 → 677 passed (+ D5 scheduler=9 + concurrency=12); phase0 anchor 0.9600 / 0.8408; roadmap D5 done, endpoint schema registry deferred, D6 optional derived panel store.
• AGENTS.md — made byte-identical to CLAUDE.md.

D5 scope (as recorded)

GlobalRateLimiter (thread-safe ticket limiter, injectable monotonic/sleep, no secret); request_with_retry(scheduler=) acquires a global slot per attempt; data_update.concurrency.max_workers default 1 (<1 rejected); run_data_update builds one scheduler only when max_workers>1, shared by all feeds; TushareCache(max_workers>1) parallelizes only the dense per-symbol fetch, then upserts/records on the main thread in deterministic plan order (successes durable, failures uncovered+retryable). Default serial; existing configs validate unchanged; no factor/alpha/portfolio/runtime/backtest math change.

Endpoint schema registry / D6 — NOT implemented

Confirmed. This is a publish + docs-sync task only; no endpoint schema registry, no D6 PanelStore append/partition / materialized derived panel store, no data-quality / research change, no further concurrency expansion.

Verification (docs PR)

• git diff --name-status main...HEAD → only AGENTS.md, CLAUDE.md, docs/data/data_layer_contracts.md (docs only).
• cmp -s AGENTS.md CLAUDE.md → identical.
• git diff --check main...HEAD → clean; merge-marker scan → NO_MERGE_MARKERS.
• stale-text scan: no 656 passed; no "D5 deferred/unimplemented" residual; 677 passed recorded; endpoint schema registry / D6 not claimed implemented.
• docs-only ⇒ full pytest not required (the D5 suite already 677 passed on feat(data): D5 opt-in bounded concurrency with a global rate limiter #51).

PR #38 untouched

Confirmed: #38 OPEN base=feat/i5c-mmp-minute-factor — not merged, closed, retargeted, or edited.