docs(sitemap-promote): seed twitter + hackernews PoC at global path

[jackwener]

Summary

Promote Sitemap Hub PoC v1.1 from local-only to global seed for two complementary sites:

• twitter (12 files, dense UI + testid-heavy + COOKIE_API) — by @opencli-user
• hackernews (10 files, SSR + no testid + structural selector) — by @opencli-质量官

Both seeds land at skills/opencli-sitemap-author/references/site-memory/{twitter,hackernews}/sitemap/ and bump frontmatter source: local → global. The two PoCs together close the cross-validation loop on the v1.1 schema (#1822) — complex SPA vs simple SSR, write-heavy vs read-heavy, modern testid vs structural sibling-traversal — both encode cleanly under the same schema and tooling.

v1.1 schema application checklist (per #1822)

Both PoC sets apply the 12-patch schema delta as a unit:

1. Form B compact YAML actions (pre / do / post / fail / recover / evidence, ~80 tokens) used as default; Form A markdown reserved for actions that genuinely need prose explanation.
2. adapter_health_update: <adapter> -> suspect|broken write directive on every adapter-primary action and workflow Fallback path — closes the read↔write loop so opencli-browser-sitemap consumer can mutate local overlay.
3. Regularized delimiters: | for failure-signal enums, || for fallback priority in do:, ; for sequential recovery steps.
4. selector_pattern first-class anchor: forms used across both PoCs include id-anchored, sibling-traversal (HN tr.athing[id="<id>"] + tr a[href^="item?id="]), data-testid (Twitter [data-testid="like"]), form-name, ARIA role. Discouraged forms (nth-child(<rank>) / single class / text-only) avoided.
5. Partial pages (_ prefix + empty url_patterns: []) used for cross-page UI primitives — Twitter _tweet_card.md references from home / profile / status.
6. 800 token hard limit respected per file — see size table below; both PoCs natural-fit without forcing.

Cross-PoC token sizes (per file, bytes ≈ token×1.5 for English-heavy content)

Site	Min	Median	Max	Total
twitter (12 files)	1.7 KB	2.4 KB	3.1 KB	29.7 KB
hackernews (10 files)	1.0 KB	1.7 KB	2.5 KB	17.5 KB

Per-file all under the 800-token hard limit. Insight from 质量官: simple sites with 1-2 actions per page naturally sit at 800-2000 tokens — don't force-split when natural. Author-side empirical guidance (3-tier <1500 / 1500-3000 / >3000) will land in a follow-up SKILL.md patch.

Trust-reality verification

Both seeds were verified end-to-end against live sites before promotion:

• twitter: 4 action chains tested (like / repost / bookmark via adapter; reply via DOM fallback); pitfalls.md records 6 verified site-specific traps (login wall / mobile UA redirect / queryId rotation / reply composer quirk / RTL caret / non-English locale).
• hackernews: 3 workflows (read-story / submit-story / upvote) tested across news / newest / item routes; pitfalls.md records SSR-specific traps (login-required-for-vote / nth-child anti-pattern / age-gated submit).

What's NOT in scope (deferred)

• SKILL.md size guidance 3-tier patch — follow-up PR (per thread agreement). Rationale: keeps this PR pure-seed for clean diff review; size guidance is author-side empirical complement to schema's hard 800 limit, deserves own description space.
• More sites — PoC stops at 2 cross-validating cases; expansion (e.g., toutiao / xhs / linkedin) waits for opencli-browser-sitemap skill to land Phase 2 features.

Test plan

• CI green
• LGTM from @opencli-质量官 (review by author of hackernews seed)
• Reviewer spot-checks frontmatter source: global on all 22 files
• Reviewer spot-checks at least one Form B action in each PoC for delimiter regularity
• Reviewer spot-checks at least one workflow Fallback path for adapter_health_update directive presence

cc @opencli-质量官 (review owner per thread)