Beagle SCM

DIFF-001: be diff: must recurse into submodules and render gitlink/pin changes

be diff:?<sha> on a commit whose only change is a submodule pin bump shows NOTHING: the gitlink pointer move (old-pin → new-pin) is not rendered and diff: never recurses into the sub, so the real work is invisible from the parent. diff: should render the gitlink change AND recurse, the way GET/POST/PATCH already do.

Input

Context

• A pure pin-bump commit diffs to nothing: the <sub> gitlink change is not emitted, and diff: never reaches the sub's content.
• Consequence: a parent commit that advances a sub by real work looks empty; review/audit of a pin bump is impossible from the parent.
• Surfaced verifying ABC-002: be diff:?29df3828 (bumps abc b8ed6277→d8656f41; the FILEErr change lives in the sub) → empty output.

Goals

• Render a submodule gitlink change as a diff entry <path> <old40>..<new40>.
• Recurse diff: into the mounted sub and emit its commit-range diff (path-prefixed under the mount), mirroring GET/POST/PATCH; --nosub suppresses.
• Keep a non-sub be diff: unchanged.
• Tests: assert (a) the gitlink line and (b) the sub's content diff for the range.

Constraints

• Localized to the diff: renderer + the sub-recursion the other verbs do.
• Pin-range diff data is graf's (DIFFREF.c); submodule mount discovery is be's job — graf must not gain mount discovery (per-store boundary).

WIP

Design decisions

• Part-b: option B chosen (be-driven relay, 2026-06-15). be already owns submodule mount discovery + child dispatch (BEProjectorSubs → BESubsHere → BERelaySub); it merely replays the parent URI verbatim per sub. The fix is diff-specific logic in be's fan-out: for diff:, rewrite each sub's child URI to diff:?<old-pin>#<new-pin> from the pin pair graf already computes for the gitlink line (part-a). The relay stays generic for every other projector.
• Why not A (graf-driven): graf would need submodule mount discovery, breaking its per-store boundary — discovery is be's job. Rejected.
• Pin source: route #3 — structural via keeper (2026-06-15). SUPERSEDES the first impl's route #1 (scraping graf's rendered <sub> <old40>..<new40> diff text — fragile, couples to render output). Pins come from keeper/SUBS.h KEEPSubsAt(tree) (per-tree gitlink+path enumeration, built on dog/git/SUBS.h SUBSu8sParse): new = subs at the target commit's tree, old = subs at the base commit's tree, joined by mount path (new-only ⇒ added/null old, old-only ⇒ deleted/null new). No capture-pipe detour, no bediff_parse_gitlinks. Endpoints MUST match graf's diff base (single sha → vs first-parent; ?a#b → a vs b).

TODOs

• Part-a: render the gitlink change (old-pin → new-pin) — diffref_collect_visit collects WALK_KIND_SUB (is_sub) and emits <path> <old40>..<new40>; test test/diff/08-sub-pin-bump-gitlink.
• Part-b (route #3): be's diff fan-out builds the per-sub pin map from KEEPSubsAt over both endpoint trees and rewrites each bumped sub's child URI to diff:?<old>#<new>; relay + path-prefix; --nosub.
• Part-b test: be diff: on a pin-bump shows the sub's content diff (test/diff/08); added-sub range (test/diff/10).

Blockers and bummers

• None — fully landed.

Outcome

• Part-a LANDED ef0430da (gitlink line; validated vs the ABC-002 bump → abc b8ed6277..d8656f41).
• Part-b LANDED 018a59c4 — route #3 structural pins via KEEPSubsAt (commit-show → first-parent; ?a#b → a/b), sha1hex pins, owned path8b paths, URIMake-built child URIs; graf untouched. 288/288 green. DIFF-001 complete.