fix(e2e-real): standalone perf gate must exit nonzero on a regression (#214 follow-up)

[VijitSingh97]

Summary

sudo bash tests/e2e-real.sh perf (the release perf gate per RELEASING.md) printed a red ✗ perf REGRESSION line but still exited 0, so any scripted check of the exit code would pass a regressing release.

• perf() now ends every path through summary "perf", mirroring how provision/verify/teardown end — summary is what converts FAIL>0 into exit 1. All five early return 0 paths (bench failure, unparseable hashrate, record-mode regression refusal, record-mode success, no-baseline) now flow through summary "perf" too.
• In all mode a perf failure is now attributed to perf (the run stops at perf's summary), not misattributed to teardown.
• The e2e-real: perf recording must judge before it writes — EPYC regressions are currently recorded, not gated #214 unit test in tests/run.sh no longer redefines the exit plumbing it is supposed to verify: the harness now extracts the real ok/bad/summary from e2e-real.sh (verbatim, same technique as the rig_lock tests) and asserts on perf()'s real exit status — nonzero on a stubbed regression, 0 on a pass. Only phase and the bench remain stubbed.
• Audited the other standalone dispatch phases for the same pattern: provision, verify, and teardown all already end in summary with no early return after a bad (verify's one early return is a clean SKIP that goes through summary "verify" first). Only perf had the bug.

Verified: the reworked test goes red against the pre-fix perf() (the two exit-code assertions fail with rc=0) and green with the fix. make test: 1332 passed, 0 failed.

Closes #267

🤖 Generated with Claude Code