Bound every skill-package read at 8 MiB

[dacharyc]

Fixes #87.

What was wrong

readFileWithCap truncated to 8 MiB only after util.SafeReadFile had already loaded the whole file, so the cap never bounded memory. And the cap only guarded token counting: the unclosed-fence check, the orphan scan (three read sites), and skillcheck's readers loaded reference files whole with no limit at all. Fixing only the token path would have left "bound memory during file reads" untrue of validate structure as a whole.

What changed

• util.SafeReadFileN(root, path, limit) applies the existing regular-file and symlink-containment checks, then reads through an io.LimitReader. It reads one byte past the limit to report truncation without a second stat, and never more.
• util.SafeReadFile now refuses files over util.MaxSkillFileBytes (8 MiB) with util.ErrFileTooLarge instead of reading them. Every existing caller is bounded through this change with no call-site edits; only the callers that should say something were touched.
• Token counts read a bounded prefix. types.TokenCount gains a Truncated flag (also in JSON as "truncated": true), and a warning names the file and the limit. The per-file hard limit still fires on the prefix count, since 8 MiB of anything is far past 25,000 tokens.
• Unclosed-fence check skips an oversized reference file with a warning rather than checking a partial read.
• Orphan scan marks an oversized file as unscanned and warns, ahead of the orphan warnings its unseen links may cause, so a false "unreferenced" report has its explanation next to it.
• SKILL.md over the limit fails to load with "reading SKILL.md: file exceeds the read limit: ... is larger than 8 MiB".
• skillcheck readers (content and contamination analysis) skip oversized files silently, as they already did on any read error; the structure warnings above cover the same file.
• README documents the limit under the structure checks.

Verification

Peak resident set on a skill with one 214 MiB reference file:

Build	Peak RSS
main	940 MB
this branch	291 MB

The remaining 291 MB is the tokenizer working on the 8 MiB prefix, which is now the ceiling regardless of file size. The same run shows all three new warnings on the file.

Tests: SafeReadFileN truncation and containment, SafeReadFile refusing a sparse file one byte over the limit and accepting one exactly at it, and one test per check for the truncated-count warning, the skipped fence check, and the unscanned-file warning ordering. Sparse fixtures keep the over-limit tests free of disk cost. go test ./... and golangci-lint run are clean.

No CHANGELOG entry, following the release-commit pattern.