Evidence method
===============
Two independent T3 dev servers were run from the specified commits, each with
T3CODE_DEV_SEED=0 and its own worktree-local database. Both used the same stdio
MCP fixture (slow-mcp.mjs), initialized after four seconds, configured as the
user-scope server ticketdesk in a scratch CLAUDE_CONFIG_DIR/.claude.json.
Scratch settings retained the existing provider connection credentials locally;
credentials and raw transcripts are deliberately not included in this upload.

The sole local source adjustment on BOTH commits was changing
DEFAULT_IDLE_TIMEOUT_MS from 30 * 60 * 1000 to 45 * 1000 in
apps/server/src/orchestration-v2/ProviderSessionManager.ts.

capture.py drives the real app through its CLI, first looking up T-1, waiting
until the original Claude process exits, then looking up T-2 in the same thread.
Arguments: --checkout CHECKOUT --config SCRATCH_CLAUDE_CONFIG_DIR
--output NEW_OUTPUT_DIRECTORY --side before|after --commit COMMIT
--model YOUR_CONFIGURED_MODEL. Start the dev server separately with the same
CLAUDE_CONFIG_DIR and T3CODE_DEV_SEED=0. Dependencies must already be installed.

The videos are live timestamped command/provider-transcript captures rendered
while the real commands run. They are not web-client screen recordings. The
JSON event files contain those same observations. The separate resumed-turn
text files preserve selected request-content excerpts from an earlier identical
run, including the exact false disconnect/reconnect notices on base.

Only synthetic ticket data is included. Runtime model identifiers, credentials,
and private machine/workspace metadata have been excluded.

AI-assisted; verified with the tests and tools listed above.