Show full-precision scores; add submissions show --no-code

[robobryce]

Two small quality-of-life fixes to the submissions views.

1. Full-precision scores in list and show

Scores were formatted with {:.4}, rounding the geomean leaderboard score to 4 decimals. That's lossy enough to make near-tied submissions indistinguishable — two submissions scoring 0.0017318... and 0.0017449... both render as 0.0017, so you can't tell which is faster from the CLI.

Switch to f64::to_string() (via a small format_score helper), which prints the shortest decimal that round-trips to the same f64 — full precision, no trailing-zero noise.

submissions list — before / after:

# before
ID       Leaderboard   File            Time                 GPU(s)   Status   Score
830213   qr_v2         submission.py   2026-06-23T12:00...   B200     done       0.0017
# after
830213   qr_v2         submission.py   2026-06-23T12:00...   B200     done       0.001731805142084383

submissions show 830213 — before / after:

# before
  - leaderboard on B200: passed (score: 0.0017) [secret]
  - leaderboard on B200: passed (score: 0.0017)
# after
  - leaderboard on B200: passed (score: 0.0017448536290123567) [secret]
  - leaderboard on B200: passed (score: 0.001731805142084383)

2. submissions show --no-code

submissions show always prints the full submission code, usually the largest part of the output. Add a --no-code flag to omit it when you only want the metadata and per-run scores. Default behavior is unchanged.

popcorn-cli submissions show <id> --no-code

Verification

Built the binary and ran both against the live API: list and show print full-precision scores; show --no-code prints metadata + runs and omits the code block (default show still prints it).

• cargo fmt --all -- --check — clean
• cargo clippy --all-targets --all-features -- -D warnings — clean
• cargo test — 31 passed (added format_score + truncate unit tests for patch coverage)

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 25.00000% with 9 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/cmd/submissions.rs	30.00%	7 Missing ⚠️
src/cmd/mod.rs	0.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[brycelelbach]

codecov/patch is failing you should fix that.

[brycelelbach]

Your PR description does not have a before and after example of submissions list only submissions shown.

[brycelelbach]

Mark/other humans: I think --no-code or something similar is important for agents to avoid context blow out. I am not sure what the default should be. The best default for agents would be to not show code by default and require a --code flag to opt-in. However, that would be a breaking change, so I opted for --no-code instead. Thoughts?

[robobryce]

Fixed in 32f4ea5. I pulled the score-rendering logic (shared by list and show) into a format_score helper and added unit tests for it (plus a truncate test). codecov/patch's target was 4.09%, and the new tests cover the bulk of the changed lines, so patch coverage should clear it now. The --no-code branch and the clap wiring in mod.rs are still exercised only by the live show path (the cmd/ layer has no unit tests today), but the meaningful new logic is now covered.

[robobryce]

Added a submissions list before/after to the description (e.g. 0.0017 → 0.001731805142084383).

[robobryce]

On the default: I agree code-by-default is the wrong default for agent use (it's the single biggest source of context blowout in show). My take, FWIW:

• This PR keeps --no-code (opt-out) to stay non-breaking, which I think is the right move for this PR.
• If we're willing to take one breaking change, I'd actually prefer default to no code, add --code to opt in — it's the better default for both humans (who usually want the metadata/scores) and agents. Could pair it with a deprecation cycle: add --code now, keep printing code by default but print a "code shown by default is deprecated, pass --code" notice to stderr, then flip the default in a later release.
• A lighter-touch alternative: keep code-by-default for humans but auto-suppress when stdout isn't a TTY (agents/pipes), which avoids a hard breaking change. Happy to go whichever way you and Mark prefer — I can adjust this PR or split the default-flip into a follow-up.