docs: correct loop-control subsystem descriptions in ARCHITECTURE.md/PRODUCT.md

[TheGreatAxios]

Summary

• docs/ARCHITECTURE.md and docs/PRODUCT.md described a loop-control subsystem (repetition detection, tool-fingerprint thrash/backstop, requireEdit, never-acted/never-edited/no-progress/repetition stop classes, hard-blocked re-dispatch, default 30-turn leaf cap) that was deleted in a recent release. Rewrote those sections to describe current behavior, verified line-by-line against the tree.
• No code changes.

What changed and why

ARCHITECTURE.md

• SubAgentDirector paragraph: removed never-acted/never-edited/no-progress/repetition/requireEdit; described the surviving stops (complete, incomplete-report/incomplete-report-stop, turn-budget only when maxTurns is finite, report-forced, cancelled) and that requireEvidence only checks reads (shell reads included via shell-evidence.ts), with no write-evidence requirement.
• Same-brief ledger: it counts dispatches and gives advisory hints (retry with higher maxTurns, then suggest a different approach after 3 same-brief turn-budget salvages) but never refuses a re-dispatch — the hard-block is gone.
• Model-family policy / main-session loop protection: removed the entire tool-fingerprint exact-period thrash detector, hard-pause, and turns-since-user-message backstop description (with its fleet-heavy exemption) — none of it exists in src/agent/director.ts anymore. Main-session loop protection is now the one-shot toolOnlyNudgeFired check-in nudge only; there is no automated stop for the primary session.
• Sub-agent stall paragraph: stalled now compares itself to turn-budget/cancelled, not to the deleted stop classes.
• Precedence paragraph: stall sits below turn-budget only (no-progress no longer exists).
• Watchdog exemption (~line 377): rewrote to say the task tool exemption is unconditional, not because the leaf is otherwise bounded — maxTurns is unbounded by default, deadlineMs is opt-in, and there is no no-progress/thrash stop.
• Removed the "hard cap 4 workers" line from the Skywalker system-prompt description — CL-7001 is removing the same claim from the actual prompt text (src/agent/prompts.ts, skywalker/package.ts) since no concurrency cap exists in code; landing both changes keeps prompt and doc consistent.

PRODUCT.md

• Removed the "hard cap 4 workers" claim (line 145) — no concurrency cap exists anywhere in src/; it survived only as prompt wording, not enforcement.
• Rewrote the worker hard-stop paragraph (line 162) to remove the "2 consecutive identical tool calls" / no-progress claim and describe the actual leaf stops (incomplete-report-stop, turn-budget only when finite, stalled, cancelled/deadlineMs).
• Rewrote the same-brief fingerprint paragraph (line 163): it's an advisory parent hint after repeated turn-budget salvage, not a refusal.

Found beyond the listed scope

• src/subagent/intervention-log.ts has a stale doc comment citing IDENTICAL_REPEAT_MIN and TURNS_SINCE_USER_MESSAGE_BACKSTOP (both deleted). Not fixed here: this PR stays docs-only, and an in-session request to fix it mid-review was blocked by the harness's own src/-write guard, which is correctly enforcing that boundary while other lanes touch src/subagent/. Flagging for a follow-up.
• ARCHITECTURE.md's "fix for four prior mechanisms" claim (fleet authority tiers section) was checked against current code and is still accurate as written — left unchanged.

Test plan

• bun run check (lint/typecheck/build/test) — all green, 5340 tests passed, 0 failures.

No CHANGELOG entry — docs-only change.

[linear-code]

CL-7003