docs: architect prompt upgrades — mandatory state-save compaction (arch-init) + PRFT at human decision points (role)

[waleedkadous]

Two architect-prompt upgrades from Waleed's asks, both docs-only.

1. State-save compaction is mandatory (arch-init skill, 4 byte-identical copies)

The save discipline had compaction as soft step-3 advice; in practice state files grew and a sibling architect appended a duplicate "How to resume" section — the ever-expanding failure mode. Now:

• Step 1 bans duplicate headings and requires deleting resolved loops (their record is the log entry).
• Step 3: compaction is part of the save, not optional polish — keep recent dated entries, collapse older ones into a one-line summary that points at the durable artifacts (PRs/issues/reviews) they covered. No hard numeric cap — the size guidance is judgment-shaped: "one screen is the right order of magnitude, judged by you." (An earlier revision of this PR had a hard ~80-line cap; removed per project convention — no hard limits in instructions.)
• Corrected a factual error: architect state files are gitignored, so there is NO git-log fallback for pruned prose — compaction must replace detail with pointers, never delete the only record.

2. PRFT at human decision points (architect role, instance + skeleton)

New subsection in Integration Review: when bringing a fix to the human for a decision (merge word, pr gate, dev-approval), present Problem / Root Cause / Fix / Testing unprompted, whatever the risk tier — root cause verified by the architect, each part tight, decision answerable without opening the diff.

Also updates the spec-1134 content pins to track the new wording. Byte-parity across the four skill trees verified (single md5).