ci: make doc-review link check case-sensitive per path segment

[junlin-luo-spacemit]

Problem

The repo keeps finding the same bug class: a reference that resolves on case-insensitive Windows/macOS but 404s on Linux CI and the web host. PR #138 fixes the current instances by hand — this PR makes the linter actually catch them, so it stops recurring.

The existing links.py was too permissive in some ways and too strict in others:

• It only compared the filename, so a wrong-case directory (Static/) slipped through.
• It ignored HTML <a href>, so anchor-style links were unchecked.
• It ignored #fragment anchors entirely, so links to headings that no longer exist went unnoticed.
• It pressured authors to rename the file rather than fix the reference — the opposite of what we want.
• product_names.py scanned the whole line, including image paths, so a correct lowercase path like static/com260_model.png was reported as a branding violation.

Changes

checks/links.py

• Verifies every path segment, not just the filename — directories are included.
• Covers HTML <a href> in addition to markdown links/images and HTML <img src>.
• Validates #fragment anchors against the target page's real heading slugs (both in-page and page.md#frag).
• Reports the corrected reference text to copy, instead of implying the file should be renamed.
• Still skips fenced code blocks and site-absolute (/...) paths.

checks/product_names.py

• Now prose-only. Strips code spans, link targets (keeping display text), HTML tags, URLs, and path/filename-like tokens before matching, so image paths are no longer flagged as branding violations.

checks/_paths.py (new)

• Shared helpers for both checks: reference parsing, per-segment case walk, apply_case_fix, and slugify/heading_ids for anchors.
• Resolves .. segments before the case walk. Without this a relative path escaped the check and produced false "missing file" reports — latent, since the production caller always passes absolute paths.

Cleanup

• Removed PyGithub from requirements.txt — declared but never imported (the agent uses requests directly).
• Replaced stale source_of_truth entries in entities/{k1,k3,p1,p1s}.yaml with comments; the paths pointed at key_stone/ and power_stone/ directories that never existed, and nothing reads them.
• Fixed glossary.yaml header comments that referenced non-existent files.

What this does not do

reviewer.py is intentionally untouched, so the job remains non-blocking. This PR only changes what the checks report — it is a correctness improvement, not a policy change. Turning findings into a merge gate is a separate decision (see below).

Verification

• All ten check modules import cleanly after the refactor.
• Reference coverage is unchanged at 671 relative references before and after the rewrite — no detection was lost.
• Repo-wide run over 94 markdown files: 0 link/heading issues (after PR docs: fix K3 CoM260 image reference casing for case-sensitive hosts #138).
• Sensitivity tested with deliberately faulted files: a 4-fault file and an 11-fault file were both fully caught, and a decoy inside a fenced code block was correctly ignored — confirming the check can actually fail.
• get_errors clean on the changed modules.

Follow-up (not in this PR)

If we want this to gate merges, three non-blocking mechanisms have to be removed (continue-on-error: true, the || true, and sys.exit(0) in reviewer.py) and the job marked required in branch protection. Worth noting that a full-repo measurement shows ~48 of 94 files currently trip at least one check, so a hard gate would fail most doc PRs until a baseline or diff-scoped enforcement is in place. Recommend starting with links + headings only, which are clean today.

Risk

Low. Tooling only, no .md files touched — meaning the doc-review.yml workflow (which triggers on paths: "**/*.md") will not even run on this PR. No runtime behavior change for the site build.