fix(#44): retry grouping — hide startup errors from real turns

[lis186]

繁中摘要

Codex session 把啟動失敗的 502/429/499 retry（0.1 秒、無 output）跟真正的 API 對話混在同一個 turn list，使用者看到 66% failure rate 但實際 100% 成功。

這個 PR 在 client 端把 retry 從 real turn 分開：

• Session card 顯示 2t 2r（2 real turns, 2 retries）而非 6t
• Turn list 隱藏 retry card（drill-down 仍可查看）
• 全 retry session 顯示 No turns — 1 failed request (504) 而非空白沉默
• Gap timing、compression detection、keyboard nav、cost efficiency 都跳過 retry

只改 client-side（2 個 production 檔 + 1 個測試檔），server 完全不動。Heuristic 在 27,219 筆真實 log 驗證 99.9% 準確。

---

Problem

Codex sessions mix startup 502/429/499 retries (0.1s, no output) with real API turns (85–249s, has output). Users see inflated failure rates and a cluttered turn list.

Expert consensus (Tufte/Norman/Charity Majors roundtable):

• Tufte: proportional ink violation — 0.1s noise gets same visual weight as 249s real turn
• Norman: conceptual model mismatch — user's "turn" ≠ system's "turn"
• Charity Majors: alert fatigue — red !http badges lose meaning when 4/6 turns show them

Solution

Heuristic — isRetry = !isHttpStatusOk(status) && !(output_tokens > 0)

Validated against 27,219 real log entries with independent adversarial verification (6 auditors + 1 judge, 13 claims cross-checked). 99.9% accuracy; 26 edge cases deferred to Phase 2.

Session card — shows real turn count + retry badge:

BEFORE: gpt-5.5 · 6t          (4 of 6 are retries — misleading)
AFTER:  gpt-5.5 · 2t 2r       (2 real turns, 2 retries)

Turn list — retry entries hidden (still in allEntries for drill-down):

BEFORE                        AFTER
#1  ● !http   0.1s   —       #1  ○ 200   85s   $0.42
#2  ● !http   0.1s   —       #2  ○ 200  249s   $1.87
#3  ● !http   0.1s   —
#4  ○ 200     85s   $0.42    (2 retries accessible via session card badge)
#5  ● !http   0.1s   —
#6  ○ 200    249s   $1.87

Empty state — all-retry sessions explain instead of silence:

No turns — 3 failed requests (502 × 2, 429)

Downstream filters fixed:

• Gap timing backward scan skips retries (no more 0.1s noise as gap baseline)
• Compression detection backward scan skips retries (no false compaction from partial billing)
• getVisibleTurnIndices excludes retries (keyboard nav skips hidden entries)
• renderCostEfficiencyPanel filter aligned with sparkline (input_tokens > 0)

Files changed (3, client-side only — server untouched)

File	Change
public/entry-rendering.js	isRetry computation, retryCount counter, allEntries flag, gap timing + compression scan exclusion, early return to skip card rendering
public/miller-columns.js	Session card retry badge (Nr), getVisibleTurnIndices filter, cost efficiency filter alignment, updateRetryEmptyState() for all-retry sessions
test/retry-grouping.test.js	15 difference tests covering classification, counters, session card, keyboard nav, gap timing, compression, Claude regression, empty state

Verification

Layer	Result
Red→green TDD	12 tests written before implementation, confirmed red on old code, green after
Adversarial audit	6 independent auditors verified heuristic counts, blind spots, blast radius; 3/13 claims corrected
Full test suite	836 tests pass (0 fail), including all e2e
Browser smoke	Real Codex session 019e9225 (1t 1r) — retry hidden, badge visible
Claude regression	Session ee89450f (521t) — zero retries, completely unaffected
Keyboard nav	ArrowUp/Down confirmed to skip retry indices
Empty state	Session 019e929d (0t 1r) — shows "No turns — 1 failed request (504)"

Known limitations (Phase 2)

• 15 entries with status=101 (WS upgrade shells + interrupted turns) bypass the heuristic — needs transport-aware classification
• 21 entries with status=200 + SSE error bypass — needs SSE event type detection
• Specimen 2 entries (6 total, input_tokens > 0 but output_tokens = 0) leak through sparkline filter

Closes #44

🤖 Generated with Claude Code