feat(memory): add native explorer tabs and ACP analysis

[phodal]

Summary

Add a global Memory workbench backed by native Rust discovery and authorized snapshot reads. The left explorer organizes personal guidance, general guidance, projects and source files; opening an item creates or activates an editor tab with its own reading position. Memory stays independent of the active Studio project.

AI analysis uses the existing ACP Agent/model settings, streamed conversation, permission controls and follow-up composer. The server rereads and checks the selected snapshot digest, fixes its source/section for the session, and cleans up the temporary working directory on close or disconnect. No native Memory write API is introduced.

Scope and evidence

• Native sources: Claude Code, Codex, Qoder and Qwen Code. Version 2 separates library identity, native binding, material role and content scope while preserving the legacy CLI projection.
• Bounded section extraction retains exact digest/line provenance. Browsing reads at most one canonical summary/index per global library, up to 8 files of 256 KiB; other source bodies are read only after selection. History/skill/support files remain source material.
• The old example Inbox no longer occupies the default Memory route. AST analysis, persistent promotion, native Memory mutation and automatic ADR/Wiki writes remain out of scope.
• Windows root validation now waits for the complete root before checking metadata; symbolic links and reparse points remain rejected. Native session slugs omit the Windows IO-only verbatim prefix, including equivalent drive and UNC forms. CLI help/JSON baselines include the new Memory command, and browser CI provisions a Rust Evidence Host.

Maintainer request; no tracker issue assigned. Specs: native sources, explorer and ACP, CI readiness. Current design: Memory ADR.

Validation

Latest local results with Node 24, including integration with current main:

Check	Result
Studio build and complete unit suite	612 passed
Complete Rust Evidence Host suite	46 passed (includes 7 Memory tests and Windows path-form equivalence)
CLI contract, native Memory bridge and doc links	20 passed
Complete Studio browser suite after main integration	106 passed / 1 skipped (includes 15 Memory tests)
Root CI suite	1746 passed / 7 environment-based skips
Shared ACP browser regressions	6 passed
Package verification	696 npm entries / 958 runtime zip entries
Real native library	2165 files / 87 sources; 1440, 1024 and 390 widths; light/dark, keyboard focus, overflow and no console/page errors
Actual Qoder ACP	Host-provided model list, complete analysis of synthetic input with line references, successful session close

Real personal Memory was not sent to AI. Browser coverage includes delayed reads, scope/origin authorization, virtualized trees, tabs and reading position, ACP model changes, visible output before completion, multi-turn continuity, and temporary-directory/control cleanup.

Final head 53d409a5d1dfced6ad157d8332a82c44add22e61: all 7 hosted checks passed. CI run covers Windows/macOS/Linux and Node 22/24; Desktop run passes all three platforms, including Rust/native suites and packaged startup smoke. Windows Memory and native session path fixes have hosted receipts. Previously failing date and notebook tests are fixed by the main commits integrated here.

Boundaries

• Standalone Studio/CLI requires a provisioned Evidence Host or BETTER_HARNESS_EVIDENCE_HOST; no JavaScript discovery fallback is shipped.
• ACP's temporary working directory is not an OS sandbox. The selected Agent retains its normal protocol capabilities, permission behavior and native history behavior.
• Installed Desktop/NSXPC Memory acceptance is separate from source/browser validation.
• Generated by Codex under maintainer direction. Changes are scoped to Memory, its native bridge, validation and related documentation. No release/version metadata changes.