scan finds no sessions when the repo's git dir lives outside the working tree (.git file / --separate-git-dir)

[bryantsai]

Symptom

Running backpass scan in a repo whose working tree has a .git file pointing at an external git dir (created with git init --separate-git-dir ~/repo.git, or by tools that keep the git dir outside a synced folder — my case is an Obsidian vault in iCloud) associates zero transcripts. Every Claude Code session for that repo lands in .backpass/scan-cache.json with tier: null, even though each transcript's cwd is exactly the repo root.

Cause

listWorktrees() in src/repo.js builds the candidate list purely from git worktree list --porcelain. For a main worktree with an external git dir, git reports the git dir path as the worktree line, not the working tree:

$ cd ~/vault && git worktree list --porcelain
worktree /Users/me/.vault.git
HEAD ...
branch refs/heads/main

So listWorktrees(root) returns ['/Users/me/.vault.git'], tier-1 association (realpath(cwd) ∈ repo.worktrees) never matches, and tier 3 only kicks in for paths that no longer exist. Setting core.worktree doesn't change git's output either.

Repro with backpass's own code:

import { listWorktrees } from "backpass/src/repo.js";
import { associate } from "backpass/src/discovery/association.js";
const root = "/Users/me/vault";           // has a .git *file*
listWorktrees(root);                       // → ['/Users/me/.vault.git']
associate({ cwd: root }, { worktrees: listWorktrees(root) }); // → null

Fix

Seed the worktree list with realpathOrSelf(root) so the working tree is always a candidate (the Set dedups the normal case). PR incoming with a regression test using git init --separate-git-dir.

Tested on backpass 0.1.22, git 2.x, macOS.

[kunchenguid]

Speaking as Kun's firstmate: confirmed on latest main (bd2b2cb / 0.1.22). Real bug: for git init --separate-git-dir / a .git file, git worktree list --porcelain reports the external git dir, so listWorktrees() never returns the working tree and tier-1 association misses every session whose cwd is the repo root.

Contract-class: restore. Unconfigured scan already promises that a live repo-root cwd associates at tier 1; the empty-list fallback in listWorktrees already intended to keep the root as a candidate. Seeding with realpathOrSelf(root) restores that promised path when git reports only the external git dir.

VISION per-rule

• Evidence is the only currency: aligns (correct association keeps real transcripts in the corpus).
• The human owns the weights: aligns (discovery only; no write / review-gate change).
• Nothing the model says is taken on faith: aligns (no staging / measurement change).
• The budget is the constraint: aligns (no always-loaded growth).
• It reads what you already have and owns nothing: aligns (local git layout; accuracy over quiet miss — a drifted/unusual store must not look like no history).
• Failure is loud and named: aligns (root-cause fix + --separate-git-dir regression shape).
• Scope: aligns (repo discovery / association only).

Tracked by PR #128 (Fixes #127). Labeled bug + ready-for-pr. No competing PR from me.