docs(spec): Phase 3 design for AI humanisation of commit history

One-liners per commit and a progress narrative per card, plus the
swappable provider layer both depend on, ported from the chat/ structure
in MEGAIFRASTRUCTURE.

Model and cost are measured rather than assumed: the full 779-commit
backfill costs about ninety cents on claude-opus-5, so model choice here
is a quality decision and not a cost one. Batch API considered and
rejected — it saves roughly forty cents and adds a job state machine.

Records the rules the non-editable preamble must guarantee, each tied to
a specific way a summary could mislead a paying client.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/superpowers/specs/2026-09-16-phase-3-ai-humanisation-design.md

@@ -219,8 +219,15 @@ methodology panel applies to the effort figures.
 6. Editing the appended prompt text cannot remove any rule in §5.1.
 7. Incrementing the prompt version marks existing summaries stale without
    rewriting them.
-8. A reverted commit reads as a reversal, verified against a real revert in
-   WFDASH-WEBAPP's history.
+8. A reverted commit reads as a reversal, and a removal reads as a removal.
+
+   Checked against the real history: WFDASH-WEBAPP contains **no `revert:`
+   commit at all**, so the reversal rule must be tested with a crafted message
+   rather than a live one. It does contain many removals — `chore(login):
+   simplify to login-only, remove register/reset tabs`, `feat(collection-confirm):
+   rewire page to use PoCard + remove Adjust Quantity` — and those are the real
+   risk: a removal summarised as an addition, or as a loss of function the client
+   did not agree to, is wrong in a way they will notice.
 9. No summary contains a date, a deadline or a completion claim.
 
 ## 12. Suggested plan split