[baron]

Summary

Fixes #972.

Claude's compact_boundary reports pre_tokens, not post-compaction occupancy. The old estimator retained that count, its active-turn observation, and historical fallback; UI projections could then substitute the whole transcript's character estimate and label unknown usage Live. Reopening a session could rebuild the same pre-compaction count.

• Clear invalidated occupancy at the boundary, retain the known window, and display an explicit Pending / Awaiting post-compaction usage state in the context pill and both sidebar surfaces. Do not substitute transcript estimates while pending; unknown rings show ?, not zero, and nil occupancy is not labelled Live.
• Append a zero-cost contextCompacted usage-ledger fence. The builder never scans across that fence, including normal hydration and reopening from history. Existing usage/billing records remain intact; legacy rows decode with no fence.
• Carry a dedicated pending flag, reconstructed from the fence and cleared by fresh occupancy. Keep the historical compaction timestamp without using it to suppress later rebuilding.
• Restrict the existing system-text recognizer to the compaction prefix, so Task notification prose containing the phrase does not invalidate parent usage.
• Replay realistic NDJSON envelopes (session_id, uuid) through ClaudeSDKNDJSONTranslator and the actual view-model stream handler. Exercise non-empty transcript analytics via update(snapshot:), manual compaction with no post-size, same-turn automatic compaction with fresh usage before result, Task-prose negatives, legacy decoding, hydration, and history reopen.

Relevance and integration overlap

Issue #972 remained open/reproducible on base ab84d47e6; no existing #972 PR was found. Draft #1458 (#1019) is functionally distinct but does have textual merge conflicts in ClaudeContextUsageEstimator.swift, verified with git merge-tree using original heads 3e92c3e and 39355a0. It was not merged into this branch.

Draft #1458 still has a simple adjacent-line conflict in ClaudeContextUsageEstimator.swift; no merge was performed. When integrating #1458, keep this PR’s unconditional marker-aware rebuild and add its configuredContextWindow(for:) capacity fallback and this PR's marker-aware replay/pending flag. This follow-up removes the old canRebuildContext gate because the persisted fence makes rebuilding safe. The integrated builder should receive existingUsage?.modelContextWindow ?? configuredContextWindow(for: session); retain both the pending-state helper and #1458's configured-window helper. Do not resolve the conflict by dropping either behavior.

Validation

All builds/tests used this worktree's conductor daemon; SwiftFormat mutation was limited to changed files.

• Original fail-first: 1 regression test, 12 expected assertion failures on the old estimator (ticket 4c0e1ee4).
• Final expanded root selection: make dev-test FILTER='ClaudeCompactionContextUsageTests|ClaudeCompatiblePluginBridgeTests|AgentSessionLinkContextLoadTests|AgentRuntimeSidebarViewModelTests' — 65 tests passed (ticket 323155f3; reconnected to the same ticket after one client timeout).
• make dev-swift-build PRODUCT=RepoPrompt: passed (ticket fb170bda).
• make dev-provider-test: all 20 tests passed (ticket 384691ff).
• Changed-files-only SwiftFormat 0.61.1; make dev-lint: passed (ticket 034425bd).
• Follow-up static Oracle review: no material findings.
• Commit and push preflights: passed, including redacted secret scans and repository guardrails.

An intermediate expanded run compiled successfully and passed 63/65 tests; the two failures correctly exposed an empty test fixture, fixed by saving an explicit non-empty canonical transcript. The corrected run above includes both hydration paths. The initial cold-build attempt before the first implementation hit XCBuild input/database I/O errors; the coordinated retry supplied the fail-first evidence.

Not run: full root suite, pr-ready, packaging/release validation, or live UI/provider/MCP smoke. Live smoke is out of scope for these lanes. No visible app launch, stop, or relaunch. Draft retained for independent adversarial verification.

Round-2 review follow-up

• Clear awaitingPostCompactionUsage when the provider changes, so pending Claude occupancy does not leak into ACP. Preserve the known window and restore the estimated transcript fallback immediately.
• One regression exercises the actual provider-switch coordinator and non-empty sidebar snapshot; it failed before the fix (ticket fd229c43).
• Updated focused selection: 66 tests passed (ae986fe8); app product build passed (2cc33f12); changed-files-only formatting and make dev-lint passed (54aea581). Provider-package sources were unchanged in this follow-up; prior 20-test package validation still applies.
• No full suite, live smoke, live provider run, app lifecycle, packaging, or release validation performed in this follow-up. Live smoke remains out of scope.