fix: http_probe HEAD uses curl no-body mode; analyze_pdf follows its dependency

[Calmingstorm]

Two pre-existing defects found during the v3.65.0 smoke test. Neither was introduced by the workspace change, and both were design-discussed with Odin before any code.

1. http_probe HEAD hung for 5s and reported a spurious error

build_http_probe_command appended -X {method} for any non-GET, so HEAD became curl -X HEAD — which sends the HEAD token but leaves libcurl expecting a response body.

Measured against a healthy server:

-X HEAD : exit=18  5.107s   "transfer closed with 741 bytes remaining"
-I      : exit=0   0.065s

HEAD now takes curl's native no-body path:

• -I, and no -X override
• the unconditional -i suppressed for HEAD — -I already routes headers to output, and combining them makes the output contract depend on the curl version
• -L kept: curl -I -L performs HEAD across the chain, matching the requested method and the existing follow-redirects contract. The final result comes from the -w values, which parse from the existing ---PROBE-RESULTS--- sentinel rather than from body termination.

HEAD is the only affected method. POST/PUT/PATCH/DELETE/OPTIONS may return zero-length bodies and curl frames those normally — verified, -X OPTIONS exits 0 in 0.067s.

Live-verified all four shapes against odin-bot.net:

case	exit	time	header blocks	status
HEAD, no redirect	0	0.07s	1	200
HEAD, redirecting (-L)	0	0.10s	2	200
HEAD, follow_redirects=false	0	0.07s	1	301
GET (control)	0	0.07s	1	200

Contract change

A nonempty body on HEAD is now rejected rather than silently dropped. Data flags plus -I make curl's method selection ambiguous, and silently discarding hid a caller mistake. An empty body is still accepted. The pre-existing test that asserted the silent drop is updated to the new contract.

Also note: the previous test_head asserted "-X HEAD" in cmd — it pinned the bug, so it is replaced rather than extended.

2. analyze_pdf was advertised everywhere and installed nowhere

PyMuPDF lives in the optional pdf extra, and packaging/postinstall.sh, scripts/incus-deploy.sh and Dockerfile all ran a bare pip install .. So the tool was advertised on every install path while its dependency was present on none of them, and calls died with No module named 'fitz'.

Both halves, per Odin's design call:

1. Gate visibility — the catalog hides analyze_pdf when the module is not importable, joining claude_code, the email tools, issue_tracker and generate_image in tool_catalog's existing backend-gating. Startup logs why it was hidden.
2. Install the extra in packaged paths — all three now install .[pdf], so an official install genuinely provides what it advertises and a rebuild does not undo a manual repair.

PyMuPDF is deliberately not promoted to a base dependency: it is heavy and native with its own platform and security-update burden. pip install . stays minimal; .[pdf] is the source opt-in.

find_spec is treated as structural availability only — it proves importability, not that the native library loads — so the handler converts any import/load failure into a clean PDF support unavailable message naming the remedy.

A trap worth calling out

TestBackendGatedVisibility asserts exact arithmetic: len(catalog) == len(EXPECTED_TOOL_ORDER) - len(GATED). The seven existing gated tools are config-gated, so that constant is deterministic. This gate is environment-gated, and fitz is absent from the dev venv but present on the live install — so a fixed constant would have passed on one machine and failed on the other. That is the local-green/CI-red failure mode from PR #239. The test now derives its expectation from actual availability and separately asserts that visibility tracks the dependency in either direction.

Verification

• Mutation-verified: restoring -X HEAD, dropping the HEAD-body rejection, removing the catalog gate, or restoring the raw import fitz each fails its test.
• Full suite run as a normal user and as root (passwordless sudo, CI-like): 7,647 / 7,644 passed.
• Gates: lint-no-new 0, types-no-new 0, coverage-no-drop findings 0, git diff --check clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_011BDkRoJv6pTMTTYrh1SPzy