feat(seo-review): score table, trend arrows, unlimited scan with batching

[AriaEdo]

Summary

Overhauls the seo-review workflow with a deterministic score, week-over-week trends, unlimited page coverage, and Claude batching for large sites.

Score & trends

• Adds a deterministic score table (overall grade 0–100 + 8 issue tallies) at the top of each weekly issue
• Adds trend arrows (▲ improved / ▼ worsened / ▬ unchanged) vs the previous weekly audit
• Previous scores are persisted in a hidden <!-- seo-scores: {...} --> comment inside the issue body — no extra storage required
• Workflow fetches the latest closed seo-audit issue and passes its body as PREVIOUS_AUDIT_BODY env

Scan coverage

• Removes MAX_PAGES cap — weekly audit now covers all 92 (and beyond) built pages. The env remains as an opt-in escape hatch when set to a positive int.
• Workflow drops the MAX_PAGES: '25' env

Claude integration

• Chunking by char budget: when the report doesn't fit in a single call, splits into batches, runs a "notes-only" prompt per batch, then a final synthesis call merges results into one unified review
• Safety fallback: if combined batch notes approach the cap, deterministic concat keeps the run robust
• Prompts now require a 💡 Fix suggestion for every issue (Critical / Improvements / Per-page notes) — output went from 16 → 33 actionable fix lines per weekly run
• max_tokens raised 2048 → 8192 (output was being silently truncated)
• Token usage logging — per-call and grand-total stdout lines for cost observability

Refactor (perf + cleanup)

• Cheerio.load() once per page; $ shared between extractMeta and extractLinksAndRefresh (was parsing twice in full mode)
• File reads parallelized via Promise.all (was sequential await loop)
• Cache repeated selector queries in extractMeta
• Drop dead siteHost plumbing
• Rename local path shadowing node:path in findRedirectChains

Verification

Smoke-tested three paths against the live dist/client (92 pages):

Path	Wall-clock	Tokens	Output
Single-call (default 640K cap)	~34s	in=21,944 out=3,923 (25,867 total)	12.7 KB
Batch + synthesis (forced 50K cap, 2 batches)	~45s	in=26,301 out=6,144 (32,445 total)	7.2 KB
Batch + fallback concat (forced 8K cap, 8 batches)	~50s	~80K combined	9.6 KB

Score for current main: 🟡 89/100 — surfaces real issues including a systemic 2-H1 bug across ~55 /handbook/** pages, a typo in /changelog/ og:title ("shippped"), and relative og:image URLs across 8 /brand/* pages.

Issue: #1348

Test plan

• Trigger workflow manually via workflow_dispatch (mode: full) and verify the issue body contains:
  • Hidden <!-- seo-scores: {...} --> comment
  • ### 📊 Score summary table with grade + 12 metric rows
  • Trend legend ("No previous audit found" on first run; arrows on subsequent runs)
  • AI review section with 💡 Fix: lines per issue
• Confirm token usage line appears in workflow logs (Token usage — N call(s): ...)
• Open a PR that modifies an src/pages/** file and verify changed-only mode still posts a sticky PR comment with the score table
• After the second weekly run, verify trend arrows compare against the previous run

[github-actions]

🔎 SEO & Meta Review

Skipped (mode: skipped-no-changes).

No CHANGED_FILES provided in changed-only mode.