Consolidate bounded pane context with GetPaneContext - #884

Summary

Implements the consolidated, bounded context operation for #838, named GetPaneContext as agreed during implementation.

• Append GetPaneContext to the existing COM protocol without changing earlier vtable slots; advertise protocol 2.3 and get_pane_context.
• Add wtcli get-pane-context, with version/capability negotiation before calling the appended method and validation of targets and capture budgets.
• Resolve the source pane once inside Terminal and return metadata plus bounded command output, or a bounded recent buffer tail when command marks are unavailable.
• Route planner, manual /fix, and error-triggered Autofix context collection through WTA -> wtcli -> the existing COM server. No direct Rust COM client and no new COM server.
• Preserve a logged, capability-gated legacy path for older servers. Ordinary request failures do not retry against another pane.
• Bound capture at the buffer-reading stage, preserve Unicode, and report truncation accurately, including the newline-lookahead edge case.
• Address review feedback by processing immutable captured text on a background thread while keeping UI references apartment-safe, and reject malformed, zero, or explicitly empty targets before COM connection.
• Include the separately authored logging change approved for this PR: global Debug logging no longer enables ACP dependency prompt-body traces unless that dependency is explicitly configured.

The Autofix event-delivery and cross-tab routing fixes were isolated and already merged in #880. They are part of this branch's main baseline, not new fixes in this PR.

Debug benchmark results

Important: this is a same-build comparison of two context-collection paths, not an A/B comparison of two separately built branches.

Both paths ran against the same locally deployed x64 Debug package built from dev/vanzue/get-context, based on main commit db609f8061f81c2eb9a4bdaf3e0666392596bce4, with the feature changes applied:

• Package: IntelligentTerminal_0.8.0.2_x64__rd9vj3e6a2mbr.
• Build layout: src\cascadia\CascadiaPackage\bin\x64\Debug\AppX.
• Both paths used the same packaged wtcli.exe and the same running Terminal/server. Deployed WTA and WindowsTerminal hashes matched the build outputs.
• Legacy: the benchmark script reproduces the serial collector in main at the baseline commit above, using the old protocol operations. It does not execute a main-built WTA binary. Explicit Autofix retains the old initial active-pane call, followed by window/tab/pane enumeration and output capture. There is no artificially added unsupported-capability probe.
• Consolidated: one wtcli get-pane-context process captures the context. Authentication and capability negotiation still happen within wtcli; one process does not mean only one raw COM call.
• Each scenario used the same stable target pane for both paths, five warmup pairs and 40 recorded pairs, alternating execution order. Percentiles use nearest rank and exclude warmups.
• The measured boundary is the sum of wtcli process startup through exit and stdout/stderr completion. It includes authentication, COM, and capture, but excludes PowerShell JSON processing, full prompt assembly, model inference, and network latency.

In these measurements, median context-collection time decreased by 47.9%-88.9%, saving approximately 80-685 ms per collection. The new path's p50 stayed at approximately 84-87 ms.

Interpretation and limits

• The main benefit is removing repeated wtcli process launches and serial queries, not parallelizing the original calls.
• Explicit Autofix's legacy request count depends on how many tabs must be visited before finding the source pane. The three fixture panes occupied different positions in the same topology. The 6/8/9 request counts and the largest speedup are scenario-specific, not universal and not solely attributable to scrollback length.
• Both paths read the same source pane and expected marker. Content/metadata stability was checked for each path across all samples. Bounded output was identical for the unmarked fixtures. Marked output was not byte-identical: the old capture returned one concatenated line, whereas the consolidated capture preserved 11 lines. Raw hashes and lengths record this difference rather than assuming identical payloads.
• This does not establish the improvement to full user-perceived AI response time, nor a full main-build-versus-feature-build comparison.
• Release measurements have not been run. Debug numbers must not be presented as Release results.
• These measurements predate the latest merged-main and review-fix commits; they retain their original build provenance and are not claimed as new measurements of the latest head.

Reproduction and artifacts

The PR includes test\e2e\Measure-PaneContext.ps1, its shared helpers, 16 hermetic selftests, and usage documentation in test\e2e\README.md.

Example, with an already-running Dev package and a stable existing pane:

$env:ITE2E_PACKAGE = 'Dev'
.\test\e2e\Measure-PaneContext.ps1 `
    -Package Dev -Configuration Debug `
    -Mode ExplicitAutofix -TargetPaneId '<source-pane-guid>' `
    -Scenario marked-short -ExpectedMarks Marked `
    -ExpectedMarker '<known-output-marker>' `
    -OutDir test\e2e\artifacts\pane-context-benchmark\debug-marked

For Planner or ManualFix, focus the expected source pane first; the script does not change focus, send input, change settings, or start/stop Terminal. Each run exports samples.csv, requests.csv, metadata.json, and summary.json, including binary hashes, source provenance, and timing definitions.

Local measurements are retained under test\e2e\artifacts\pane-context-benchmark\, with the combined table in debug-comparison.csv. These ignored local artifacts are not committed or uploaded; the measured summary is reproduced above.

Coverage and validation

• Full WTA unit suite including the logging change at 7c3bdf7c2: 2,085 passed, 1 ignored. The following review-fix commit does not change Rust.
• Fresh C++ protocol-parsing tests after the review fixes: 5 passed; bounded ControlCore buffer-tail test: 1 passed (existing binary).
• Benchmark hermetic selftests: 16 passed.
• Review-fix Debug package built and safely deployed. Packaged WTA, WindowsTerminal, wtcli, and TerminalApp hashes matched the build outputs; user settings were preserved.
• All nine paired live benchmark scenarios completed, each with 40 recorded old/new pairs.
• Added unit assertions that planner and both Autofix source-selection modes each make exactly one context request, and that ordinary failures never enter the legacy fallback.
• Preserved the seven packaged PaneContext E2E cases: marked capture, unmarked fallback, explicit tab/split routing, missing/closed targets, metadata-only capture, Unicode/truncation bounds, and source resolution from focused agent panes.
• Fresh six-suite Dev run: 30 passed, 3 failed, 0 skipped across PaneContext, ShellIntegration, AutofixParser, AutofixPane, AutofixRouting, and SessionHookRouting. Both WSL cases ran and passed.
• Fixed the new empty-target test's PowerShell argument binding and reran the complete PaneContext suite: 7/7 passed, including malformed/zero/empty CLI diagnostics. The first run also exposed an existing concurrent session-index append race, isolated in Prevent concurrent helper session-index corruption #885 rather than included in this optimization or hidden by the successful replay.
• Replayed the complete five-case Autofix card/reject Describe: the first attempt was blocked before the cases by a Copilot ACP initialization timeout; a retry passed 5/5, with Esc dismissal taking 7.61 seconds under the unchanged 12-second timeout. The original intermittent Esc failure was not reproduced and its root cause is unproven; no speculative product change or relaxed assertion was made.
• All 33 distinct cases now have passing results across the combined run and targeted replays, including both WSL cases. This is not a claim that the first combined invocation passed 33/33.
• Fresh PaneContext results mark C300-C306 complete through both full and incremental report generators. Overlaying the successful Autofix replay leaves no failed checklist items in the combined incremental report. Autofix routing remains C299, preserving main's hook cases C296-C298. Original failure artifacts are retained separately.

Remaining validation

• Finish native rebuild and targeted validation for the suppressed review finding that context columns were always zero; the local fix reads the authoritative viewport width.
• Track Prevent concurrent helper session-index corruption #885 for the independent session-index race fix; retain the intermittent Esc/startup evidence for follow-up if reproduced.
• Obtain a latest-head Copilot review with no unanswered findings and a passing actual check-spelling check.
• Record Release p50/p95 measurements.