feat: support context length shorthand (64k, 128k)

[devangpratap]

What

Add shorthand notation for the --context-length / -c flag across all
commands (whichllm, plan, upgrade, run). Users can now write
whichllm -c 64k instead of whichllm -c 64000.

Supported formats: plain integers (4096), k/K suffix (64k = 64000),
m/M suffix (1m = 1000000), fractional (1.5k = 1500). Case-insensitive.

Why

Closes #59. Converting context lengths to raw integers is annoying when
the values are commonly expressed as 64k, 128k, etc.

Testing

• Tests pass (uv run pytest, 202/202)
• New tests added (12 tests in test_utils.py covering valid inputs,
  edge cases, and rejection of bad values)
• Tested on real hardware (if hardware-related)

Notes

• ContextLengthType is a custom click.ParamType in utils.py that
  handles parsing before typer sees it
• All 4 --context-length options in cli.py wired via
  click_type=CONTEXT_LENGTH
• Help text updated to show shorthand examples
• Plain integer input is fully backwards compatible

[Andyyyy64]

Thanks, this is the implementation I’d like to go with for #59.

Could you make one small adjustment before we merge? 64k currently parses as 64,000, but for context windows users may expect 64k to mean 65,536. I think we should either switch k to 1024-based parsing or explicitly document that this is decimal shorthand.

Could you also carry over the CLI docs coverage from #60? After that this looks good to merge.

[devangpratap]

Thanks for the feedback! I've pushed the changes:

1. Switched k to 1024-based parsing — 64k now parses as 65536 instead of 64000. Same for m (1024²) and b (1024³). This matches how LLM context windows are conventionally expressed.
2. Added CLI docs coverage from feat: support context length shorthand #60 — updated docs/cli.md with shorthand mentions in the option descriptions and added shorthand examples for all four commands (main, plan, upgrade, run).

All 202 tests pass, ruff clean.

[Andyyyy64]

Sorry for the slow follow-up, work has been busy on my end.

This looks good now. The 1024-based shorthand behavior matches how people usually talk about context windows, and the CLI docs coverage is in place.