fix(capacity): require power evidence for M5 promotion

[tcconnally]

Summary

• Require independent AC/lid/sleep power evidence before an M5 capacity profile can promote a cell mode.
• Carry the power_sleep_evidence gate into normalized matrix profiles and fail closed when it is missing, warning, or blocked.
• Add a regression test for the reproduced case where an 8-cell analysis with a blocked power gate was selected as PASS.

Evidence

• Before the fix, a direct repro on main 4475e0ff8f9207f28f92f0067d9d0ba8b739094d returned selected_cells=8 even when the analysis contained power_sleep_evidence: BLOCK.
• The new regression test was run RED before implementation, then GREEN after the fix.
• python3 tests/run_all.py: PASS on the exact candidate.
• python3 -m compileall -q hyperwall scripts tests: PASS.
• git diff --check: PASS.

Scope

This is a follow-up contract fix for #79. It does not claim native M5 acceptance, does not promote any cell count, and does not close #79. The Mac/native evidence gate remains pending.

Refs #79