Add a harness-neutral reasoning-effort setting to the spec loop

[justinmclean]

Summary

Add a SPEC_LOOP_EFFORT setting to the spec loop that takes low,
medium or high, and translate that level into whichever equivalent flag
the selected agent CLI understands. Unset means unset: no flag is added.

Background

tools/spec-loop/loop.sh drives a headless agent CLI through repeated
build, plan, update and consolidate passes. Several of those CLIs expose a
per-invocation control over how hard the model thinks, under different flag
names and different level vocabularies. The loop cannot express that today,
so a mechanical consolidate pass and a hard build pass run at the same
setting.

The framework never binds to one vendor's flag ([PRINCIPLES.md §9](https://github.com/apache/airflow-steward/blob/main/PRINCIPLES.md)),
so low/medium/high is the framework-facing vocabulary and each
harness maps it to its own flag. Where a CLI offers more levels, map high
to its highest sensible value and record that choice in the spec table.
spec_loop_launch_agent already does exactly this kind of translation for
output format, turning one request into --output-format, --format json,
--json, or nothing depending on the harness.

Where to look

• tools/spec-loop/lib.sh — spec_loop_launch_agent. The output-format
  handling inside each per-harness branch is the pattern to copy.
• tools/spec-loop/loop.sh — the SPEC_LOOP_MODEL block shows how a
  setting is read, printed in the startup banner and documented.
• Related: tools/spec-loop/tests/test_runner_fixtures.sh — asserts on the
  argv the runner builds, including a kiro case asserting --model is
  absent. New assertions go here.

Acceptance criteria

• SPEC_LOOP_EFFORT=low|medium|high reaches spec_loop_launch_agent
  and is mapped to the right flag for each harness whose CLI has one,
  taking the flag name from that CLI's own --help.
• A harness with no such knob adds nothing and prints no warning, the
  treatment kiro already gets for --model.
• Any other value fails at startup with a message naming the accepted
  values, before the first iteration launches.
• With the variable unset, the argv for every harness is unchanged.
• test_runner_fixtures.sh covers set-and-mapped,
  set-but-unsupported-harness, and unset, and passes.
• README.md and the harness table in specs/spec-loop-runner.md
  document the setting and the per-harness mapping.

Estimated effort

~2-3 hours for someone new to the codebase. Most of it is reading each
agent CLI's --help to fill in the mapping; the shell change is small and
has a pattern to copy in the same two files.

Getting started

[How to contribute](https://github.com/apache/airflow-steward/blob/main/CONTRIBUTING.md#your-first-contribution)

---

This issue was drafted with the help of an AI-assisted tool and reviewed by a Magpie maintainer before posting. If anything here is unclear or looks wrong, say so on the issue: a real person is reading.

[Shaurya2k06]

I'd be happy to take this up!
cc @justinmclean

[justinmclean]

@Shaurya2k06 thanks if you have any questions just ask

[Shaurya2k06]

I've given this one a shot, first time contributing here, let me know if the above PR is a good solution to the issue! I've used Grok 4.5(High) to help me write this (explicitly mentioned with PR) , code review is strictly my own.

cc @justinmclean