Addresses review follow-ups on this PR. Four changes in one commit:

1. cmd/bd/where.go: findOriginalBeadsDir's BEADS_DIR fast-path used to
   return "" unconditionally when BEADS_DIR was set but the env'd
   directory had no redirect file. Since gastownhall#3230 rebinds BEADS_DIR to the
   post-redirect target during command startup, that shortcut silently
   hid the redirect source from `bd where` output — `bd where` inside a
   redirecting worktree rendered without its "(via redirect from ...)"
   footnote. Fall through to the filesystem walk instead; keep the
   fast-path's positive branch so a user-set BEADS_DIR pointing at a
   redirecting directory is still returned directly.

   Regression tests in cmd/bd/where_test.go cover both directions:
   BEADS_DIR set + no redirect → walk finds the redirect source;
   BEADS_DIR set + has redirect → env value returned directly.

2. internal/beads/beads_inherited_artifacts_test.go: close the codecov
   gap in hasBeadsDatabase by exercising the dolt/ and embeddeddolt/
   directory branches (previously only *.db was covered). Add a
   table-driven runWorktreeSeparateDBPreservedTest helper so the three
   variants share setup.

3. internal/beads/beads_inherited_artifacts_test.go: new test
   TestFindBeadsDir_WorktreeNoDatabaseAnywhereFallsBackToLocal exercises
   the lenient-escape-hatch path in FindBeadsDir step 3b: when neither
   the worktree nor the shared fallback has a database, return the
   worktree-local .beads/ so a fresh `bd init` can bootstrap.

4. internal/beads/beads.go findDatabaseInTree: apply the same
   CanonicalizePath treatment as FindBeadsDir so the gitRoot boundary
   comparison is robust against /var → /private/var and
   case-insensitive-filesystem mismatches. Not a bug fix (this path
   requires an actual *.db to match, so inherited metadata can't
   masquerade as a real project), but consistency with the sibling
   resolver closes a latent pitfall.

…#3230)

* fix: rebind target workspace for db commands

* fix: isolate rebinding from caller workspace state

* fix: restore no-db target selection semantics

* fix: honor no-db target precedence consistently

* test: restore db flag state in context rebinding tests

…3425)

* fix(worktree): don't prefer inherited .beads/ over the shared fallback

`bd worktree create --help` promises that worktrees "automatically share
the same beads database as the main repository via git common directory
discovery — no manual redirect configuration needed." That promise is
broken when the parent repo commits its `.beads/` artifacts (metadata.json,
config.yaml, issues.jsonl, etc.) — a common setup because these files are
the portable/durable record of the beads project. `git worktree add` checks
out those tracked files into the worktree's `.beads/`, but the gitignored
`.beads/dolt/` data directory is not checked out. `FindBeadsDir` sees the
worktree-local `.beads/` with its inherited metadata.json and returns it,
instead of falling back to the shared `.beads/` with the actual database.
bd then tries to auto-start a Dolt server against an empty data directory,
which fails with:

    failed to open database: database "<proj>" not found on Dolt server
    at 127.0.0.1:<port>

...and every subsequent `bd` invocation from inside the worktree errors.

Two root causes, both fixed here:

1. `FindBeadsDir` step 2 walks up from CWD with a boundary check at
   `git.GetRepoRoot()`. On macOS, `os.Getwd()` returns unresolved
   `/var/folders/...` while `git rev-parse --show-toplevel` returns the
   canonical `/private/var/folders/...`. The `dir == walkBoundary`
   comparison silently never matches; the walk overshoots the worktree
   root, finds the inherited `.beads/` at the worktree root, and returns
   it — never reaching the worktree-fallback logic in step 3.

   Fix: canonicalize both `cwd` and `walkBoundary` (and the step 4
   `extendedRoot`) before comparison.

2. `FindBeadsDir` step 3b accepted the worktree-local `.beads/` as
   separate-DB if `hasBeadsProjectFiles` returned true. That check matches
   on metadata.json alone, so tracked-artifact dirs pass. The shared-DB
   fallback (3c) was never reached.

   Fix: add `hasBeadsDatabase` (strict — requires dolt/, embeddeddolt/, or
   *.db). Step 3b now requires a real database for the separate-DB verdict.
   If the worktree-local .beads/ lacks a database but the shared fallback
   has one, fall through to 3c. The lenient check is preserved when no
   fallback exists with a real database (so `bd init` in a fresh worktree
   can still bootstrap).

Tests: two new regression tests under `internal/beads/` —
TestFindBeadsDir_WorktreeWithInheritedArtifacts reproduces the bug (fails
on unpatched main) and TestFindBeadsDir_WorktreeSeparateDBPreservesLocal
guards against regressing genuine separate-DB mode.

* fix(where): fall through BEADS_DIR shortcut when env has no redirect

Addresses review follow-ups on this PR. Four changes in one commit:

1. cmd/bd/where.go: findOriginalBeadsDir's BEADS_DIR fast-path used to
   return "" unconditionally when BEADS_DIR was set but the env'd
   directory had no redirect file. Since #3230 rebinds BEADS_DIR to the
   post-redirect target during command startup, that shortcut silently
   hid the redirect source from `bd where` output — `bd where` inside a
   redirecting worktree rendered without its "(via redirect from ...)"
   footnote. Fall through to the filesystem walk instead; keep the
   fast-path's positive branch so a user-set BEADS_DIR pointing at a
   redirecting directory is still returned directly.

   Regression tests in cmd/bd/where_test.go cover both directions:
   BEADS_DIR set + no redirect → walk finds the redirect source;
   BEADS_DIR set + has redirect → env value returned directly.

2. internal/beads/beads_inherited_artifacts_test.go: close the codecov
   gap in hasBeadsDatabase by exercising the dolt/ and embeddeddolt/
   directory branches (previously only *.db was covered). Add a
   table-driven runWorktreeSeparateDBPreservedTest helper so the three
   variants share setup.

3. internal/beads/beads_inherited_artifacts_test.go: new test
   TestFindBeadsDir_WorktreeNoDatabaseAnywhereFallsBackToLocal exercises
   the lenient-escape-hatch path in FindBeadsDir step 3b: when neither
   the worktree nor the shared fallback has a database, return the
   worktree-local .beads/ so a fresh `bd init` can bootstrap.

4. internal/beads/beads.go findDatabaseInTree: apply the same
   CanonicalizePath treatment as FindBeadsDir so the gitRoot boundary
   comparison is robust against /var → /private/var and
   case-insensitive-filesystem mismatches. Not a bug fix (this path
   requires an actual *.db to match, so inherited metadata can't
   masquerade as a real project), but consistency with the sibling
   resolver closes a latent pitfall.