Beagle SCM

GET-023: be get '?#<sha>...' (trailing dots) silently resolves to the PARENT, not the sha

A ref with trailing dots — ?#<sha>... — resolves to the commit's PARENT instead of <sha> (or a clean error). The ... is read as range/ellipsis syntax and the resolver lands on the wrong commit SILENTLY. It bit the BLAME-006b land worker: a be get '?#3dc19b0a...' (meant to check out 3dc19b0a) reverted the worktree to the parent f6bea1c5 and dropped a merge hunk, forcing a hand-recovery. A trailing-dots ref must resolve to <sha> or fail loudly — never silently pick the parent. Method: Issues.

Input

Context

• Repro: be get '?#<sha>...' (literal trailing ./../...) → the worktree cur becomes the sha's PARENT, not <sha>. Observed be get '?#3dc19b0a...' → f6bea1c5 (3dc19b0a's parent). No error; exit 0.
• Likely cause: the #fragment <sha>... is parsed as a git-style range (A...B symmetric-difference / A... against an implicit endpoint), and the resolver picks the merge-base / parent endpoint instead of treating the leading 40-hex as the target. Resolution path: REFSResolve (keeper/REFS.c:467) / keeper/RESOLVE.c (the #hex → 40-hex decode). A 40-hex followed by ... should not be a range with an empty/implicit side.
• Severity: silent wrong-checkout is dangerous — it reverted real work (caught only because the worker verified the tree). Pairs with the GET-021 stale-worktree class (a wrong cur with no warning).

Goals

• SAFEGUARD (the core fix): the ref resolver CLASSIFIES the target as (valid sha | valid hashlet-prefix | known branch). If it matches NONE of those, it REFUSES with a clear error — it MUST NEVER silently fall through to a parent / merge-base / nearest-prefix / no-op resolution. ?#<sha>... is one instance: the trailing ... makes it neither a valid sha nor a branch → reject, don't guess.
• be get '?#<sha>...' then either resolves to <sha> (if trailing dots are explicitly stripped as a tolerated form) OR errors — but NEVER lands on the parent.
• A genuine A...B / A..B range, if supported, resolves to the documented endpoints; a malformed/one-sided range errors rather than guessing.

Constraints

• Fix in the ref-resolution / URI-fragment parse (keeper/REFS.c / RESOLVE.c / the canon query parse), not by special-casing be get. Don't regress valid hashlet / branch / ?#<sha> resolution. Add table cases: ?#<sha>, ?#<sha>., ?#<sha>..., plus any real range form.

WIP

Design decisions

• ROOT CAUSE (2026-06-16): be get '?#<sha>' carries an EMPTY query + a #fragment, but sniff/GET.c::SNIFFGetURI (~:2067) only read the QUERY — the empty ? matched the trunk row and the #fragment was SILENTLY DROPPED (landing on the trunk tip, not the sha). A trailing-dots ?#<sha>... additionally mis-decoded via WHIFFHexHashlet60 (ignored the non-hex tail → wrong object). [The trailing-dots*query* form `?<sha>...` was already rejected on trunk by `REFSResolveURI`'s`HEXu8sValid` gate — the bug is the *fragment* form.]
• FIX: a classify-or-reject arm at sniff/GET.c:2069 — if (has_q && $empty(query) && !$empty(fragment)) classify the fragment as full-sha (DOGIsFullSha) or hashlet (DOGIsHashlet) → GETCheckout, else REFUSE (SNIFFFAIL); plus the shared-resolver guard keeper/KEEP.c:1908 (KEEPResolveTree, KEEPFAIL). Reused existing error codes (no hand-rolled ron60).
• REFINED (maintainer): the first cut added a frag_is_spoilt_hex helper in graf/LOG.c with a hexrun>=6 && <len "botched-sha" heuristic — a MISNOMER (a fragment is valid hex or not; the heuristic also false-positived a #msg-search starting with ≥6 hex). REMOVED it: GRAFResolveTip now does the clean binary — KEEPResolveRef resolves it (sha/hashlet/branch) → object selector; else the only legit non-object form is a #<N>/#Ln COUNT (HUNKu8sFragSplit != 0, load-bearing so counts aren't refused); anything else returns the resolver's miss (refuse). So log:?#<sha>... → RESLVNONE (no silent trunk-log), #msg-search is a separate verb path (graf_head_msg_search), counts still resolve+cap.

TODOs

• Repro-first: be get '?#<sha>...' lands on the parent (RED); table over ?#<sha> / . / .. / ....
• Root-cause the range/ellipsis interpretation in REFSResolve/RESOLVE.c.
• Fix: resolve to <sha> or error; never silent-parent. Test green.

Blockers and bummers

• Shared resolver — verify all verbs (get/head/diff/blame) get the corrected behaviour, not just get.

Outcome

• LANDED d6f35358 (FF child of 453c55c2; 307/307): the classify-or-reject safeguard + the frag_is_spoilt_hex removal + the source URI now built via the composer (URIutf8Feed off a query-only uri, call()-checked at sniff/GET.c:2095, replacing the hand-rolled ?+feed). Reconciled clean onto trunk (disjoint from PTR-008/SUBS-020/SUBS-022).
• (history) FIX-READY ($HOME/todo/GET-023, base b71eba2a): classify-or- reject arm in sniff/GET.c (the empty-query+#fragment get/head path) + shared-resolver guards in graf/LOG.c/keeper/KEEP.c (covers head/diff/blame too). RED→GREEN proved with a trunk-only sniff binary (SNIFFfragref: ?#<sha> was silently landing on the trunk tip). Reused SNIFFFAIL/KEEPFAIL. Files: sniff/GET.c, graf/LOG.c, keeper/KEEP.c, new sniff/test/fragref.sh (+ CMake), keeper/test/RESOLVEURI.c (regression doc). 305/305; no regression in valid sha/hashlet/branch resolution. Awaiting gate.