feat(fileshare): the skill catches up with the tool

[sauterbe]

Why

xentral_fileshare gained three actions in xentral/agent-os (#150 windowed read + find, #151 update). The skill still listed all three under "what this tool cannot do" — which is worse than silence: an agent that believes a large file is unreadable reports failure to the user instead of windowing it.

What changed

• Actions: read documents offset/limit and what the window response reports; find and update are added.
• "A big file: window it, search it, or profile it" replaces "when read will not give you the file". Three routes — window when you know where to look, find when you don't, import_analyze when a figure is wanted (still the only honest source for a count, even on a small table) — plus the duty to pass on "more follows" and "stopped at max_hits" rather than papering over them. Only a binary over the cap has no route left.
• New "Overwriting" section, mirroring what the server enforces: your own uploads are yours to correct; a user's file means uploading a new one beside it unless they asked for that file by name (confirmed=true); an inbound email attachment never is. Plus the rules that aren't about the file — an overwrite is its own step, anchors must be unique, and you cannot replace the body of a file you could not read whole.
• Reference: two new recipes (locate something in a file too big to read; fix one line) and four new pitfalls, each one a way to be confidently wrong that the new actions make possible.
• Frontmatter description covers the new capabilities, so the capability catalogue and the plugin describe the tool correctly.

The remaining limits table is now short and true: no folders, no versions, no Office extraction, no binary edit.

Verification

• Assembled through the real backend loader: 14 566 chars, frontmatter stripped, reference folded in, and the new material present (offset, find, confirmed=true, the inbox rule, the new recipe, the "stopped at" duty).
• python scripts/validate_library.py → OK: 337 library items valid against their schemas.
• SKILL.md 182 lines, reference 103 — both well inside the 500-line guideline.

After merge

Tag, then make bump-library in the backend. The two agent-os PRs can merge independently; until the bump, the vendored skill understates the tool (it does not misstate it in a way that breaks anything — it just tells the agent to give up early, which is the bug being fixed here).