fix reasoning-effort carriers for codex and agy, which both reject --effort

[atomantic]

Summary

~effort=<level> was passed to every reviewer CLI as --effort <level>. That flag is valid for only claude and grok — codex and agy both reject it and exit non-zero before the review starts, so the reviewer's slot in the merge gate held a launch failure instead of a verdict.

Verified against the installed CLIs:

invocation	result
codex --effort max …	✗ error: unexpected argument '--effort' found (codex-cli 0.149.1 — no --effort top-level, on review, or on exec)
agy --effort max	✗ invalid --effort "max" (valid: low, medium, high)
agy --model "Gemini 3.5 Flash (High)" --effort low	✗ --effort is not supported for model …
agy --model "Gemini 3.5 Flash" --effort high	✗ --effort is not supported for model …
codex -c model_reasoning_effort=max … review --base main	✓
agy --model gemini-3.5-flash-low	✓ reports low

• codex carries effort as a config value: -c model_reasoning_effort=<level>.
• agy treats --effort as mutually exclusive with --model, which this loop always pins. Since the built-in default (Gemini 3.5 Flash (High)) is itself a variant, agy~effort= failed 100% of the time — at every level, not just ones agy lacks. agy encodes effort as a model variant, so it's resolved by picking from the agy models roster at run time. Nothing is hardcoded: the roster and level names change between agy releases.

The deeper fix

Effort was assigned generically and then clobbered per agent, so the fallthrough for any CLI without a branch was --effort — and that fallthrough is exactly what broke both. It's now a case over REVIEW_AGENT that fails closed: an unrecognized agent gets no flag and keeps prompt-advisory effort, because a wrong flag isn't a weaker review, it's a non-zero exit. Pre-flight went 55 → 30 lines and the ordering dependency disappeared.

A new effort-carrier table replaces the same rule restated across six places (Step 8, the 565-word "Pinning" paragraph, two copies in rpr.md, the README, the rationale bullets).

Test plan

• npm test — 391 pass, 0 fail.
• New table-driven contract test pins each agent's carrier, asserts no invocation passes a carrier its CLI rejects, and asserts the default is no flag. A new reviewer CLI adds a row rather than a test.
• Command forms above were run against the real codex and agy binaries, not inferred from --help.

Note

v3.33.1 touched this same area and added the line "names all three effort carriers" — two of those three were still wrong. The carrier table plus the fail-closed default is what stops the next one from being wrong silently.