tui: add debug-flag duplicate-widget diagnostics for issue #426

[SHYXIN]

What

Implements the temporary diagnostics suggested in issue #426 (assistant response rendered twice in a long TUI transcript). This is Path A — instrumentation only, no behavior change. It does not fix the underlying race (that is Path B, deferred per the issue's "keep open for additional reports" guidance).

Why

#426 reports a duplicate assistant row in sessions exceeding the 200-item mounted window, with structured (thinking + text) finalization. The durable JSONL and event pipeline contain a single copy, so the symptom is presentation-only. The issue's "Suggested next steps" ask for temporary diagnostics that:

• assert/log when multiple visible widgets represent the same canonical assistant item
• record assistant / ChatItem / mounted-widget identities, transcript window bounds, and active streaming widget identity
• capture whether a full redraw, structured finalization, window shift, resize, or duplicate end event preceded the symptom

Changes

• TranscriptView._redraw() now takes a reason argument labeling the trigger: window_shift, state_update, append_page_to_latest, append_overscan_shift, structured_finalization.
• Gated by env var TAU_DEBUG_TRANSCRIPT_DUP=1 (read once at import). When set, a post-refresh check (_schedule_duplicate_diagnostic) inspects the mounted children and logs a WARNING on tau.tui.transcript_dup whenever two or more visible widgets wrap the same canonical ChatItem (by id(item)), recording window bounds, item/role/widget details, the active streaming item id, and the trigger stack.
• The WARNING is also written to a file; the path is printed at launch:
  [tau] transcript duplicate diagnostics enabled -> <temp>/tau_transcript_dup.log
• No effect unless the env var is set — production behavior is unchanged.
• Added dev-notes/transcript-duplicate-diagnostics.md documenting the diagnostic and how to use it.

How to use

TAU_DEBUG_TRANSCRIPT_DUP=1 uv run tau chat <session-id>
# at launch: [tau] transcript duplicate diagnostics enabled -> <temp>/tau_transcript_dup.log
# after a session, check:
grep "tau.tui.transcript_dup" "$TEMP/tau_transcript_dup.log"

In normal operation the logger stays silent (no false positives). If a duplicate is observed, capture the full warning (including reason) and attach it to #426.

Verification

• ruff clean, mypy clean.
• TUI test suite passes; the 6 failures on Windows are pre-existing path-separator issues in unrelated sidebar/export tests (reproduced on clean main), not caused by this change.
• A forced-duplicate check confirms the detector fires with the expected fields.

Note

This keeps the issue open for real-world reports. If async pruning is confirmed as the cause from collected logs, a serialized/awaited DOM reconciliation (Path B) can follow without conflicting with this instrumentation.

[SHYXIN]

Closing in favor of #632, which provides the actual fix (Path B) for the duplicate-rendering bug described in #426. This PR was the temporary Path A diagnostics (env-gated duplicate detection); with #632 landing, the instrumentation is no longer needed. Issue #426 can stay open until #632 merges.