Scrub stop-policy commentary: retracted claims and threshold contradictions

[TheGreatAxios]

Closes CL-6789.

Diff is comments-only (verified with git diff -U0 filtered to non-comment lines: empty).

• Removed the fabricated session/run scan numbers ("358-session/428-run", then "428 runs") from the TURNS_SINCE_USER_MESSAGE_BACKSTOP justification comment. The retraction statement is kept, but the specifics that were retracted are no longer restated.
• No replacement justification invented for the 100 threshold — the comment states plainly it's a judgment call and points at src/subagent/intervention-log.ts (measured value logged beside each threshold) as the way to settle it going forward, rather than at another guess.
• Verified the identical-repeat threshold contradiction the ticket described (leaf repeat limit vs. director-level IDENTICAL_REPEAT_MIN) is already resolved on main: both are 5 (leaf raised from 2 to 5 in a prior commit), and the existing comment on DEFAULT_SUBAGENT_REPEAT_LIMIT already documents that it mirrors IDENTICAL_REPEAT_MIN and why — no further edit needed there.
• Checked src/agent/model-family-policy.ts for the same class of issue; found no retracted claims or contradicted constants there, so it is unchanged.

Not touched (out of scope for a comments-only change): THRASH_PARENT_HINT and isThrashSubAgentReport/appendThrashParentHint in stop-policy.ts still describe a "progressive thrash (re-read pressure)" stop, but that stop path was deleted and no ForcedStopReason ever produces a "progressive thrash" summary anymore, so this is dead code with a stale user-facing string — worth a follow-up, but fixing it means editing a string constant, not a comment.

[linear-code]

CL-6789