feat(resume): bare path resume — cross-harness session picker

[od0]

Closes #110. Top of the picker-performance stack (#156 → #157 → #159 → this) — one squashed feature commit whose tree is byte-identical to the build validated in dogfooding.

path resume with no argument opens a recency-first cross-harness session picker: the newest 100 sessions from every installed harness, ranked cwd-matches-first then newest, hydrated in milliseconds regardless of history size. Picking a session derives it (write-through cache, mirroring share), runs the harness picker (source pre-selected), projects, and execs.

Surface

Invocation	Behavior
path resume <input>	Unchanged.
path resume (TTY)	Recency picker → derive (cache fast-path) → harness picker → project → exec.
path resume (no TTY)	no input provided and no TTY for interactive selection; …
path resume --harness X	--harness stays the target, validated before the picker; harness picker skipped.
path resume --from Y	Session picker filtered to source harness Y (filter-aware empty message).
path resume --project P	Full sweep filtered to project P (its matches may be arbitrarily old).

A tail row — … N older sessions — load everything — runs the full sweep on demand.

How the picker stays fast

Codex candidates rank stat-only by mtime; each of the top 100 hydrates from the listing cache on a stamp hit (instant, message count included) or via one streaming read that's cached thereafter. Claude/opencode use the cache-backed collectors. Measured end-to-end on a 3.9 GB codex tree + 455-session claude project: first run ~2.3 s, then 0.03 s warm.

Design notes

• Cache policy deviates from the issue sketch (which proposed opt-in --cache): write-through by default mirrors share, keeps --no-cache meaning one thing everywhere, and feeds the freshness fast path plus later share/query/sync.
• A scoped-to-current-project-first design was built and reverted (history + spec addendum retained): it hid other projects behind an extra step, when recency + cwd-first ranking already surfaces the right rows.
• New SessionPicker seam (mirrors ExecStrategy) makes the whole pipeline headless-testable — FixedPicker (with multi-round sequences) + RecordingExec drive 12 bare-mode integration tests with zero TTY.
• toolpath-codex 0.6.3 adds peek_metadata (O(1) head+tail, count-free) and factors the shared HeadHunt used by both metadata paths.

Rebase notes

• ResumeArgs derives Default and all test constructors use ..Default::default() — feat(resume): path resume --remote <ssh-url> — resume a session on a remote host #145/feat(resume): remote persistence & transport picker (--persist / --via) #146's appended fields stop breaking every constructor. New flags sit mid-struct; no ExecStrategy/RealExec/fuzzy.rs changes.
• Projection goes exclusively through project_into_harness, so Plugin /path:resume + /path:link-pr; fixes that make projected Claude sessions resumable #150's ClaudeProjection change is inherited cleanly.
• Version slots: path-cli 0.17.0 here; path query: parallel per-file execution (path-cli 0.17.0) #138/feat(resume): path resume --remote <ssh-url> — resume a session on a remote host #145 claim the same — whoever lands second renumbers.

Validated by heavy dogfooding on real multi-gigabyte trees (this PR's picker is how its own review sessions get resumed).

[github-actions]

🔍 Preview deployed: https://2dc0742a.toolpath.pages.dev

[od0]

Dogfooding note: the first bare path resume on a large ~/.codex/sessions tree stalled ~80 s in the pre-picker gather. Two follow-ups landed: a "Gathering sessions..." stderr hint on this branch (7e28f07), and the actual fix — streaming codex metadata — split out as #156 since it equally benefits share and p list codex. Not a hard dependency; this PR works without it, just slowly on multi-GB codex trees.