fix(usage): OpenAI percent normalization for low values (#137)

[quangdang46]

What

jcode usage (and the inline /usage overlay) was reporting ChatGPT
subscriptions as fully exhausted (100% used, exhausted bar) even when
the live wham/usage endpoint reported only a few percent consumed.

This addresses issue #137: #137

Root cause

Both src/usage/openai_helpers.rs and the legacy src/usage_openai.rs
had a normalize_ratio_value helper that tried to auto-detect ratio-vs-
percent inputs using raw > 1.0:

if raw > 1.0 {
    (raw / 100.0).clamp(0.0, 1.0)
} else {
    raw.clamp(0.0, 1.0)   // bug: maps 1.0 -> 1.0 (100% used)
}

The OpenAI wham/usage endpoint unambiguously reports used_percent
in [0, 100]. A real low-usage payload (used_percent: 1) hit the
else branch and was clamped to 1.0, rendering as fully exhausted.
Existing tests only exercised values like 25.0, 50.0, 75.0,
100.0 so the bug never surfaced.

Changes

• src/usage/openai_helpers.rs: drop the heuristic; divide by 100
  unconditionally. Made normalize_ratio_value pub(super) so the
  test module can call it directly. Added a docstring explaining why
  the heuristic was removed.
• src/usage_openai.rs: same fix in the legacy duplicate.
• src/usage/tests.rs: two new regression tests:
  • test_normalize_ratio_value_treats_low_integer_values_as_percent
    locks the contract across 0, 1, 5, 50, 100, >100,
    negative, and NaN inputs.
  • test_parse_openai_usage_payload_reports_low_percentages_correctly
    feeds a real-shaped wham/usage payload with used_percent: 1 in
    the secondary window and asserts 7-day window == 1.0.

Tests

$ cargo test -p jcode --lib usage::tests::
test result: ok. 38 passed; 0 failed; 0 ignored

Both new tests pass; all pre-existing usage tests still pass.

Acceptance criteria

• Live wham/usage payloads with used_percent in [0, 100] no
  longer collapse low values to 1.0 ratio — src/usage/openai_helpers.rs:15-23
• Same fix applied to legacy src/usage_openai.rs:15-23
• Regression test for the helper locking the contract — src/usage/tests.rs:589
• Regression test for the full parse path with a realistic payload — src/usage/tests.rs:608
• All existing usage tests continue to pass

Notes

• Ports upstream PR Fix OpenAI usage percent normalization for low values 1jehuang/jcode#178 by @ktmyname.
• Cargo.lock version-bump hunk from upstream skipped — this fork tracks
  its own version in Cargo.toml.
• A pre-existing clippy::question_mark warning in
  crates/jcode-tui-mermaid/src/mermaid_cache_render.rs:169 is unrelated
  to this change (reproducible on master).