车东 Che, Dong

Docs: CSP nonce + strict-dynamic strategy (#158) + v4.9 PLANDOC entry

- docs/01-PRODUCT.md: new §3.2 CSP Strategy — current state, target state,
  transition plan (3-phase), implementation note for nonce sharing
- docs/05-PLAN.md: v4.9 CSP nonce + strict-dynamic entry

Co-Authored-By: Claude Code deepseek-v4-pro <noreply@taotoken.net>

fix: --status query excludes NOT_FOUND entries from emoji counts

Failed enhancements are now cached as status='not_found' to prevent
retries. The --status count query must filter status='found' to avoid
counting failures as completed enhancements.

Co-Authored-By: Claude <noreply@taotoken.net>

fix: cache NOT_FOUND on LLM failure to prevent repeated retries

Pages that fail enhancement (LLM returns empty) were not cached,
causing them to be retried every batch run, wasting consecutive
failure quota and killing processes prematurely.

Now failed pages are cached as status='not_found', so the next
run's annotation check correctly skips them. Use --rebuild to
force retry cached failures.

Co-Authored-By: Claude <noreply@taotoken.net>

fix: preg_match_all called before empty() check in splitHtmlIntoSections

The heading regex was defined but preg_match_all was placed AFTER the
first empty($matches) guard, so $matches was always undefined and the
function always returned a single chunk. Also fixed stale variable
references ($tag → $headingText, $positions → $matches).

Co-Authored-By: Claude <noreply@taotoken.net>

fix: man page heading detection in splitHtmlIntoSections

Man pages use <a id="section-NAME"><b>SECTION</b> anchors, not <h1> tags.
Previous regex required only ALL_CAPS <b> content, producing 378 false
positives. Fixed to require <a id="section-" or <a id="sub-" prefix.

Also wired chunked path into batch-enhance.php HTML phase.

Co-Authored-By: Claude <noreply@taotoken.net>

fix: wire chunked enhancement into batch-enhance.php HTML path

batch-enhance.php called callLLM() directly with full raw HTML,
bypassing enhanceManPageChunked(). Now pages > PHPMAN_ENHANCE_CHUNK_THRESHOLD
(200KB) are automatically split and enhanced in chunks.

Co-Authored-By: Claude <noreply@taotoken.net>

feat: chunked LLM enhancement for large man pages (>200KB)

Pages exceeding PHPMAN_ENHANCE_CHUNK_THRESHOLD (default 200KB) are
automatically split into sections at h1/h2 heading boundaries, each
chunk independently enhanced via callLLM(), then reassembled.

Split strategy (failsafe cascade):
1. h1 section boundaries (NAME, SYNOPSIS, DESCRIPTION, etc.)
2. h2 sub-section boundaries within oversized sections
3. Paragraph/block boundaries (</p>, <br>, </li>, </dd>, etc.)
4. Sentence boundaries ($. + ' ') — last resort, never mid-sentence

Key guarantees:
- Never splits inside <pre>, <table>, <dl> blocks
- Parent heading preserved in every chunk for LLM context
- Failed chunks fall back to original HTML (graceful degradation)
- Each chunk capped at PHPMAN_ENHANCE_MAX_CHARS (32KB)

Co-Authored-By: Claude <noreply@taotoken.net>

docs: add architectural paradox — data sources are already CLI tools

The man/perldoc/info/ri/pydoc3 data sources already have CLI interfaces.
phpMan's web layer (HTML rendering, format negotiation, emoji enhance)
is redundant for LLM Agent use cases. The only irreplaceable value:
TLDR cheatsheets + cross-mode FTS5 search.

Co-Authored-By: Claude <noreply@taotoken.net>

docs: strategic review — phpMan value in the LLM era (GEO/MCP/SEO analysis)

Key findings from production access log:
- AI bots prefer MCP/markdown over HTML (78-86% structured format choice)
- MCP is GPTBot's #1 format (437 > markdown 413 > html 297)
- Traditional SEO failed — Googlebot only 23 phpMan requests
- LLM emoji enhancement cost: ~¥2,170 total (), marginal value
- phpMan's strategic value shifted: human web viewer → AI agent API

Co-Authored-By: Claude <noreply@taotoken.net>

fix: deterministic PID file naming for multi-mode parallel batch-enhance

Problem: restarting a batch process didn't clean up the old one
because PID file paths weren't derived from the same option combination.
Manual 'rm pidfile && nohup ...' bypassed the startup guard, creating
duplicate processes for the same mode.

Changes:
- buildPidFilePath(mode, format) — canonical path for any option combo
- 15 unique PID files: 5 modes × 3 formats (both/html/md)
- --force flag: auto-kills existing process with same PID file
- --yes also enables auto-cleanup (batch/cron friendly)
- --restart + --mode=man auto-finds /tmp/bm.pid (no need --pid-file)
- Old error message now suggests --force as an option

PID file naming:
  --mode=man                → /tmp/bm.pid
  --mode=man --format=html  → /tmp/bm_html.pid
  --mode=ri  --format=md    → /tmp/br_md.pid
  --pid-file=/custom.pid    → /custom.pid (explicit wins)

Co-Authored-By: Claude <noreply@taotoken.net>