Put -n on the ranker the list already uses

[devcodes9]

The bug behind the bug

Step 2 started as "add a session id to `-n` and stop printing ANSI into a pipe". Both are real, but they are symptoms. -n and the interactive list were two different search paths:

	list	-n (before)
ranking	rank_sessions: BM25 over title / first prompt / body, recency and resume-count boost	rank_matches: all(t in line.lower()), sorted by date
stemming	yes (migration ≈ migrate)	no
typos	subsequence tier for rare terms	no
unit	one row per session	one row per message
SDK-spawned runs	demoted below your own sessions	not distinguished

Demonstrating it on my own corpus:

$ agsearch -n "heldout evluation"
No matches.

$ agsearch _filter 0 "heldout evluation"     # what the TUI runs
ef702a40…  2026-08-24  cc  1/2  agsearch  Agsearch held-out eval blind spot

The typo tolerance the README's demo GIF sells does not exist on -n.

This is not a corner. uvx agsearch -n "stripe tax id" is the README's quick start, the first command a new user runs. print_matches is also the fallback when fzf is missing. So the surface most people meet first was the unranked one, and every ranking improvement the eval measures never reached it.

A test pinned the old behaviour as intentional (test_n_mode_keeps_raw_substring_and), so this was a deliberate simplification once, not drift. It stopped being defensible when -n became the surface scripts and coding agents use.

Change

-n calls rank_sessions. rank_matches and fmt_row are deleted rather than left for the next caller to pick up by accident.

Each result is one session, led by its id, with the matching line indented under it:

ef702a40-7ed3-429c-92a4-959a19859545  2026-08-24  cc   2/2  agsearch   Agsearch held-out eval blind spot
    …65% of the held-out queries target auto sessions that can never rank 1…

The id leads because it is the only field another program needs. A hit you cannot open is a hit you cannot use, which is why the two additions below come with it:

• agsearch read <session-id> prints a whole conversation without resuming it. The code was already there as the TUI's Ctrl-O, reachable only as an internal subcommand.
• Colour is chosen at the edge, in one _emit seam. A terminal gets colour; a pipe, or NO_COLOR, gets text. That includes escape sequences quoted inside a transcript, and ones a snippet cut in half. The fzf paths pass color=True explicitly, because their pager is a pipe but their reader is human.

Cost

-n now builds the same per-session index the list does: ~0.7s per run on a 727-session corpus, up from ~0.25s. Accepted rather than optimised. The TUI pays the same cost at startup, it is a one-shot CLI call, and caching sessions.tsv against the index would be cache-invalidation work this change does not need.

Verification

Ranking is untouched, and the eval confirms it:

                         gold@1  gold-tol  goldMRR |   held@1  heldMRR |  sec/query
main                      0.733     0.800    0.795 |    0.490    0.618 |     0.0727
candidate                 0.733     0.800    0.795 |    0.490    0.618 |     0.0749

103 tests pass. Also checked by hand: _filter still emits its 4-field fzf rows, _transcript keeps colour through the pager, NO_COLOR is honoured on a tty, and read exits 1 on an unknown id.

New tests in tests/test_noninteractive_rows.py pin the contract: one entry per session however many messages match, the id leads the line, no ANSI in a pipe, a typo still finds the session.

Why now

Groundwork for #46. A skill built on the old -n would have inherited the weaker ranker and had no way to open what it found.