Project .claude/agents partially discovered in git-worktree sessions (still broken in 2.1.241; #25816 closed but not fixed)

Summary: In a session rooted in a git worktree, only a subset of the project's .claude/agents/*.md files register as agent types. This is the behavior #25816 described (closed 2026-02-18); it still reproduces on 2.1.241, with new evidence that the failure is a partial scan rather than a path-resolution miss.

Environment: Claude Code 2.1.241, Windows 11 Pro (win32), session rooted in a git worktree (D:\Work\daScript-dasweb, a worktree of D:\Work\daScript). The agent definitions are checked into the repo, so the worktree has its own copy of .claude/agents/.

Observed (one session, 2026-08-23):

• .claude/agents/ in the worktree holds 11 agent files: agent_babysitter, analysis-bot, dragon, flashlight, placement-auditor, rescue-bot, rescue-sweep-bot, review-md-auditor, spartan, style-hygiene-auditor, targeted-reviewer, tdd-auditor.
• The session-start registry contained only 6 of them: dragon, flashlight, placement-auditor, rescue-bot, spartan, style-hygiene-auditor. Absent: tdd-auditor, review-md-auditor, targeted-reviewer, analysis-bot, rescue-sweep-bot.
• Mid-session, the registry updated dynamically: a notification announced analysis-bot and rescue-sweep-bot as newly available and rescue-bot as removed. So hot-loading works for some files — yet tdd-auditor, review-md-auditor, and targeted-reviewer never appeared at any point in a session lasting many hours.
• All the missing files are valid: they load fine in sessions rooted in the main checkout, and a general-purpose agent instructed to read the definition file as its charter executes the same brief without problems (our standing fallback).
• The rejects are silent — nothing indicates a file was skipped or why (same gap as #5099).

Expected: every valid .claude/agents/*.md in the project root of the session (worktree or not) registers as an agent type, or the skip is reported with a reason.

Why this isn't just #25816 reopened with the same words: #25816 framed it as worktree agents not being discovered at all. What we see repeatedly (multiple worktrees, multiple sessions, several weeks) is a partial load — always a subset, not the same subset across sessions — while dynamic mid-session updates prove the watcher sees the directory. That points at the initial scan (or a race in it) rather than path resolution.

Repro sketch:

1. Repo with ~10 valid agent files in .claude/agents/, checked in.
2. git worktree add ../wt, start a Claude Code session in ../wt.
3. Compare the available-agents listing against the directory contents.