When all blockers are resolved, 'td show' was printing both '[all closed]'
as a prefix AND '[closed]' after each blocker ID. The prefix makes the
individual markers redundant.

Now:
- All closed: '[all closed] id1, id2' (no per-ID markers)
- All open: 'id1, id2' (no brackets)
- Mixed: 'open1, open2, closed1 [closed], closed2 [closed]'

Change TaskId serialization from the raw 26-char ULID to the short
td-XXXXXXX display form. This affects all JSON output from every
command that emits task identifiers: create, show, list, update,
done, reopen, dep, label, rm, log, next, ready, search, and doctor.

The root cause was TaskId's serde Serialize impl using
#[serde(transparent)] which forwarded the inner ULID string.
Replace it with a custom impl that emits short(). Additionally,
several commands built JSON manually via .as_str() or stored the
full ULID in plain Strings for human output — fix all of those.

Export retains full ULIDs via a dedicated Task::to_export_value()
method so that import round-tripping continues to work (import
needs exact ULIDs for CRDT key fidelity).

Update one test in cli_io.rs that was feeding show --json output
into import — the correct data source for import is export, which
is the designed round-trip path.

When both peers run 'td sync' without a project selected (no binding,
no --project, no TD_PROJECT), instead of bailing with 'both peers are
in bootstrap mode', enumerate all local projects on both sides, compute
the intersection by project_id, and sync each shared project.

Protocol change: replace the Bootstrap handshake (for the 'no project
selected' case) with a new SyncAll variant carrying a manifest of every
local project (name, id, version_vector). Both sides compute the
intersection by project_id, agree on an ordering (sort by project_id),
then loop over the shared set exchanging deltas per-project exactly like
the existing single-project Sync path.

Fallback cases handled:
- SyncAll peer + Sync peer: bootstrap from the peer that has a project
  selected, identical to the old Bootstrap path.
- SyncAll + Bootstrap (old td): bail with an upgrade hint.
- No intersection: friendly no-op message.
- Both have zero projects: no intersection, friendly no-op.
- Name mismatch for same project_id: sync anyway, print a warning.

Also adds:
- db::list_projects_in(root) - root-explicit helper used by build_local_manifest
- db::Store::project_id() - exposes read_project_id_from_doc via Store
- db::bootstrap_sync_at(data_root, cwd, project, delta) - HOME-free
  bootstrap used from async sync code where HOME may vary by peer

Tests: sync_all_exchanges_shared_projects, sync_all_no_intersection_is_noop

Add src/cmd/project.rs with handlers for all five ProjectAction variants
(init, bind, unbind, delete, list). Wire dispatch in cmd/mod.rs.

Remove the now-superseded src/cmd/init.rs, use.rs, and projects.rs modules.
Update the two error strings in cmd/sync.rs that still referenced td init.

Replace the flat Init, Use, and Projects top-level commands with a single
Project { action: ProjectAction } variant and a ProjectAction enum covering
Init, Bind, Unbind, Delete, and List.

Add unbind_project() and delete_project() to the db layer. Update the three
error strings that still referenced the old command names.

The Loro migration stripped the verbose breakdown down to just an ID and
final score. Restore the three pieces that were lost:

- Active scoring mode printed once before the per-task list
- Expanded formula with actual values substituted, branching on mode:
    impact: (downstream + 1.00) × priority / effort^0.25 = score
    effort: (downstream × 0.25 + 1.00) × priority / effort² = score
- Unblocks count: N tasks (M directly), with correct singular/plural

All data was already present in ScoredTask; only the rendering in
src/cmd/next.rs needed updating.

Update the two cli_next verbose tests whose assertions reflected the old
'score:' label rather than the restored output. Add two new tests
covering the singular 'task' grammar and transitive unblocks counts,
both of which were untested before.

short() and display_id() now emit td-XXXXXXX instead of bare suffixes.
resolve_task_id() strips the prefix on input so both forms are accepted.

next extracted task IDs as raw ULID strings for the scoring engine
then displayed those same strings. rm did the same with unblocked_ids
in its warning message. Both now go through TaskId::Display or
display_id() at render time.

Add display_id() to TaskId for contexts where a raw &str needs
shortening without constructing a TaskId.