Fix Herdr Sidebar session linkage and context trends

[lis186]

繁中簡述

這個 PR 修正並完善 Herdr Sidebar 的 session observability：恢復 native session linkage、保留啟動時的 agent status，並讓 context% 趨勢在 session 歷史超出 index tail 或只剩 contextless subagent row 時仍能正確顯示。

目前 Sidebar 會：

• 對 native session 使用精確、受限的 targeted repair，而不是把整個 index 掃進熱路徑
• 從 targeted import 保存最多 64 筆 display-oriented context samples
• 沿用 aggregate 的 beta1m 作為 context denominator provenance
• 只有在至少 3 筆有效樣本且變化達 8 個百分點時顯示 ↑ 或 ↓
• 維持固定、右對齊的 context trend viewport，避免百分比因寬度變化消失

Problem

The Sidebar intentionally reads only the last 4 MiB of index.ndjson. That bounded read can contain a matching contextless subagent row while the main conversation's context samples are older in the same session:

Full index
  [older main context samples] .............. [contextless subagent row]
       outside the 4 MiB tail                         inside the tail
                                                        |
                                                        v
                                               old behavior: matched, but `?`

The session was technically linked, so the old repair guard did not run. Separately, coarse context buckets could make a materially changing context percentage look flat, and imported rows did not always retain the authoritative 1M-window fact.

What changed

Session identity and status

• Use exact native session evidence for Sidebar linkage, with fail-closed handling for ambiguous ownership.
• Preserve native working/idle state during startup refresh.
• Keep the bounded working refresh loop alive while a pane is natively working.

Context history recovery

bounded Sidebar refresh
          |
          | matching row has incomplete context history
          v
detached targeted import
          |
          +--> exact linkage evidence
          +--> <= 64 display-oriented context samples
          v
fixed Sidebar trend + current context%

• Detect both a contextless matched tail and an aggregate whose known history exceeds the visible tail.
• Persist the bounded sample ring with the existing repair state.
• Migrate old completed repair state once so it can acquire samples without repeated imports.
• Keep request/response bodies out of the Sidebar repair cache.

Rendering and provenance

• Sort samples by received time before calculating the trend.
• Treat unknown context usage as unknown rather than zero; explicit zero remains valid only when marked as known.
• Use aggregate beta1m when available, so a 1M context window is not downgraded to the 200K fallback.
• Add guarded ↑/↓ direction cues while keeping the newest sample right-anchored.
• Keep cache and context labels distinguishable in narrow Sidebar rows.

Verification

• Herdr plugin suite: 249/249 passed
• Importer and targeted-import suites: 52/52 passed
• Full suite with an isolated CCXRAY_HOME and export disabled: 2395/2395 passed
• git diff --check: passed
• JavaScript syntax checks for all touched runtime modules: passed
• Manual Herdr Sidebar verification rendered a directional context indicator (49%↓)

Scope and limits

• The normal Sidebar refresh remains bounded; the complete index walk happens only in the detached, targeted repair worker.
• Direction cues are intentionally conservative: fewer than 3 known samples or a move smaller than 8 percentage points produces no arrow.
• The repair cache stores at most 64 samples per targeted session and is used only for Sidebar display recovery.