feat(gain): add median ratio, concentration and above-ceiling to summary - #3511

Closes #3507.

rtk gain reports totals and averages only. Command output sizes are long-tailed, so a mean describes almost none of the population.

On my install — 10,376 commands, 90-day window — the headline reads 96.5% saved, while the median command is 0.97x and 53% come back byte-for-byte unchanged. Both are true; only the first was visible.

What it adds

Three lines in the summary block, using the existing print_kpi helper:

Median ratio:      0.97x (53% unchanged)
Concentration:     top 8 commands = 92.1% of total savings
Above ceiling:     396 / 10376 commands

• Median ratio — median of output_tokens / input_tokens, skipping rows with zero input. Empty input reports 1.0 (unchanged) rather than 0.0, so "no data" never reads as "perfect compression".
• Concentration — share of total savings held by the largest 8 commands, with the count actually used so it stays honest on small datasets. This exists because a single recursive grep can dominate the total: on my data one command is 69% of it.
• Above ceiling — commands whose raw output exceeded CONTEXT_CEILING_TOKENS (7500, ~30k chars at ~4 chars/token), roughly where agent CLIs truncate. Makes the tail visible without changing anything that is stored or how savings are computed.

Scope

Additive only — 149 insertions, 0 deletions. No existing output line changes, no stored value changes, no new config key or CLI flag. GainSummary gains five fields, appended; nothing reordered.

Verification

• cargo fmt --check clean
• cargo clippy clean
• cargo test — 2585 passed, 9 of them new

The new tests cover the empty set, a single value, even and odd counts for the median, zero savings, fewer commands than the top-N window, and the single-outlier shape the concentration metric exists to expose.

Numbers were cross-checked against an independent implementation reading the same database; median, unchanged percentage, concentration and above-ceiling count all agree.

One thing worth flagging

Two tests fail on develop before this branch — hooks::rewrite_cmd::tests::unattestable_passthrough::{test_plain_command_still_rewrites, test_fd_dup_redirect_still_rewrites}. I confirmed they fail identically with this change stashed, so they are pre-existing and unrelated, but CI will be red until they are addressed. Happy to open a separate issue if that is useful.

Note on process

The display code was drafted with an AI assistant and the statistics implementation, tests, and verification are mine; every line was reviewed and built locally before pushing. Flagging it since the project seems to work with agents openly.