docs: ai-memory allowlist capture was not in force on this host (correction #144) - #988

…e working check

The GPT read of #988 found two P2s. The prescribed `ai-memory hook --check-capture` cannot run
without --event, --agent and --server-url; the doc now gives the documented form
(recipes/README.md:418) and what to expect for enrolled and excluded directories, plus the
inverse. And "lane worktrees carry none and are not captured" was not exact: a worktree below
the enrolled checkout, or one that receives the marker through .worktreeinclude, is enrolled by
design (recipes/README.md, "Worktrees and the capture marker"). The paragraph now states that
design, that read snapshots and blind lanes must never carry or inherit the marker (the read kit
aborts on any ancestor marker), and the measured state: 2 of 404 worktrees carry a marker, none
nested under the main checkout.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…read's P3 citations

The GPT micro of 77e95b7 (P2) found no regression for the two corrected statements, so the
correction rule (AGENTS.md, Corrections) was not met. Two doc-contract tests fail on the
first version's text (8690d10: 2 failures) and pass at this head:
- the documented check carries hook --event --agent --server-url --check-capture --data-dir
  (HookArgs, ai-memory v2.6.0 cli.rs:2957-2966; the bare form exits 2);
- enrollment follows recipes/README.md "Worktrees and the capture marker" (marker walk to
  $HOME, marker.rs:938-1025; .worktreeinclude), without the false "lane worktrees carry none".
The independent Claude read's P3s: the per-event read is hook.rs:720-726 and the v2.6.3
reader sits unchanged at :1383-1399; the read kit strips markers, then aborts on a remaining
or ancestor marker; the trading-research bank is cited (foundation-finalization :70,157)
instead of the unsourced "exposure tags"; the check lines are recipes/README.md:419-421.

Co-Authored-By: Claude Opus 5.5 <[email protected]>