Pass through to gh with a -- separator, delegating to gh --attach when it can do the job

[drogers0]

Implements the passthrough proposed in #58: everything left of -- is ours, everything right of it is a gh invocation.

gh image diagram.pdf -- pr create --title "Design" --body-file design.md
gh image shot.png    -- issue comment 42 --body-file note.md
gh image shot.png    -- pr create --fill

What happens, in plain terms

Two routes exist. You never pick one — the tool works out which can do the job. Either way you type one command and the file ends up in the body.

What you run	What happens	Why
A PNG or MP4 on create, with --body/--body-file	gh-image uploads, rewrites the body, hands the finished body to gh	On a create there is nothing to inspect afterwards, so we never risk a second PR
A PNG or MP4 on create, without a body flag (--fill, editor)	Handed straight to gh --attach	--fill means gh writes the body, so there is nothing for us to rewrite
A PDF, zip, log or any other type, on comment/edit	gh is tried first, refuses the type, gh-image takes over	Upstream only takes 9 image and video extensions
Any file on a repo you only have read access to, on comment/edit	gh is tried first, refuses, gh-image takes over	Upstream needs write access; our route does not
Any file, on a gh too old to have --attach	gh-image does the whole job	We check the actual gh on your machine, per command
A file on create with no body flag and no --attach support	Refused before anything uploads	Nothing could carry the reference, so nothing is wasted

Where the file lands in the body:

Your body	Result
Mentions the file — 	That reference is rewritten to the uploaded URL, in place
Does not mention it	The reference is appended at the end

Path spelling does not matter: ./shot.png, shot.png and docs/../shot.png all match the same file, the way gh matches it.

One case stops rather than guessing: if gh fails partway through a comment but has already posted it, we do not post a second one — we stop and name the files that may be missing.

Routing, precisely

• Probe. gh <subcommand> --help is checked for --attach once per run — per command, not per binary, so a build carrying the flag on only some commands routes correctly. gh resolves via GH_PATH, then PATH.
• create — delegation and fallback never mix, so the double-create hazard from upstream's partial-failure behaviour (UploadAndAttach writes when at least one asset uploaded) cannot arise: body flag → our route with no --attach; no body flag → delegate verbatim and never fall back; no body flag and no --attach support → refuse before uploading.
• comment — delegate first; on a non-zero exit check the target for a comment by the viewer since launch (with a clock-skew allowance). Something posted → stop and report. Nothing posted → our route.
• edit — delegate first; on failure fall back without a verification query, since re-execing an edit overwrites the body deterministically.

This adopts the refinement discussion on the issue with one change to the second comment's rule: create delegates exactly when fallback is impossible (no body flag), rather than never. Every upstream refusal class — extension, size, token type, permission, host — exits before any side effect, verified against cli/cli trunk, so -- pr create --fill with an eligible file keeps working instead of being refused.

Fallback rewriter (internal/passthrough)

goldmark walk of inline image and link nodes. Destinations naming an uploaded file are replaced by byte range, preserving titles and angle-bracket wrappers, skipping code spans and fences. Both sides resolve to an absolute path before matching, as upstream's rewriter does. Reference-style links error with a request for an inline link. Files the body never references are appended using the existing per-type markdown — image embed, bare video URL, [name](url) otherwise. About 280 lines against upstream's ~810, which is an internal package and not importable.

Only upstream surface knowledge taken on: the six subcommands plus their cobra aliases (pr new, issue new), the four body-flag spellings, --repo/-R for targeting the fallback upload, and --title (whose value can begin with a dash). User-defined gh aliases are refused with a clear message rather than guessed at, as is --attach right of --, which would otherwise be honoured on one route and dropped on the other.

Behaviour notes

• stdout carries what gh prints, with gh's exit code; route notes and diagnostics go to stderr. The no--- mode is unchanged.
• --body-file - is buffered before delegation so the fallback can replay it.
• Breaking: gh image -- -dashed-file.png (escape-dash) is gone — -- now always means passthrough. Use gh image ./-dashed-file.png.

Testing

Probe, exec and comment-verification are injected boundaries alongside the existing newUploader/newDownloader, so both routes are exercised with no live gh and no network: routing per kind × probe × body-flag, stdin replay across delegate-then-fallback, argv building, and table-driven rewriter cases covering inline, titled, angle-bracket and escaped destinations, path spellings, code spans and fences, and per-type appends.

gofmt, go vet, go build, go test ./... all clean. The lint job is pinned to the module's Go version in a separate commit — golangci-lint v2.12.2 cannot typecheck the Go 1.27 standard library, which fails main today as well.

Closes #58