# Development after verified Windows NSIS 3.7.3

This is an **unreleased working-tree increment**. The starting 146 recorded source hashes and NSIS digest matched before changes. The large pre-existing dirty tree was inspected and preserved. The installer remains exactly 316,758,944 bytes, SHA256 `d09d0f9418d94059e801fc43a38241314757c686e5060e26596a806ba56ae7c1`. No new installer, version promotion, commit, staging, push, publishing or deployment was performed. The production embedding model/ranking configuration and frozen 69-query dataset are unchanged.

This report describes the engineering increment before the subsequent board-guide UI fix. See `15-board-guide-ui.md` and the latest `verification.json` for that fix and its newer frontend/browser evidence. The existing `production-workflow-development.json` remains evidence of the earlier UI build.

1. **Modified files.** The recorded engineering-scope path list and before/after hashes are in `continuation-changes.json`, compared with `verification-nsis-3.7.3.json`, not with the earlier Git baseline. Files absent from that baseline are labelled as added to the recorded scope; this does not establish when every pre-existing untracked file was created. Main code: `intelligence/semantic.rs`, `sdk_process.rs`, existing SDK resolver/command registration, `CIntelligencePanel.tsx`, `CodeEditor.tsx`, `App.tsx`; typed Block converter/schema/fixtures; provider evidence transport, role checks, atomic JSON checkpoints and repair regressions. New review tooling and documentation/evidence are included; generated rule/coverage inventories were refreshed. Existing unrelated Git changes are not attributed to this increment.

2. **Implemented features.** Optional local C definition/reference navigation uses exact ESP-IDF CDB and trusted SDK GCC, unsaved buffer snapshots, request-scoped process trees, workspace exclusion, cancellation/timeouts and labelled syntax fallback. No shared semantic cache or remote index exists. The typed normalized Block subset expands to nine verified source IDs; the exact schema is inspectable offline through the existing converter tool. Tool evidence is wrapped as data, reserved role markers cannot promote user objects, and pre-replacement JSON failures preserve original recovery state. Native per-attempt repair review gates remain authoritative.

3. **Executed automated checks.** Frontend **170 tests / 18 files**; Rust **188 library + 9 example = 197 tests**; Python review validator **4 tests**. Six real native semantic-adapter queries across two generated projects pass, including unsaved definitions and project isolation. Two new actual ESP-IDF builds pass. The development production-React/CSP citation/Keep probe passes using explicit provider/native fixtures. Release/policy/diff checks and exact source hashes are recorded in `verification.json`. The tests are not a certification of every failure mode or real-model behavior.

4. **Retrieval quality comparison.** No ranker/model was retuned or rerun. Stratified analysis of the previous measured runs is in `retrieval-stratified-review.json` and 08-evaluation.md. Overall Recall@5 remains lexical **0.3478**, bundled MiniLM **0.3623**, routed MiniLM **0.3623**, E5 **0.4275**, multilingual MiniLM **0.3913**. Thai recall is **0 / .05 / .05 / .20 / .10** respectively; exact API is 1.0 on only nine probes. Reports retain MRR/NDCG/latency, process RAM, index/model bytes and actual preserved installer bytes. Hardware is a declared 33-case subset overlapping languages. There is no new measured relevance improvement, candidate installer-size measurement or human-independent accuracy claim. The new human review packet is blank; the validator rejects unreviewed/incomplete submissions.

5. **Real LLM evaluation.** **NOT_EXECUTED.** Loopback availability checks found no running model, and the user confirmed none is available. No paid-provider authorization was given; no inference requests or paid credential reads occurred. Retrieval-answer/citation/code-edit/repair quality remains null, not fabricated. The exact actual-Vibe-Coder evaluation protocol and per-patch user approvals are described in 13-independent-evaluation.md. Embedding inference, deterministic fixtures, browser stubs and clangd are not LLM evaluations.

6. **Firmware compilation.** `arithmetic_if` on `kidbright32_v13` and `nested_conditions` on `kidbright32_rev31` compile/link using installed ESP-IDF **5.5.4**, target **esp32**, no Flash or downloads. Source/firmware/CDB hashes are in `block-expansion-compilation.json`. Current regenerated main C for the two original fixtures is byte-identical to earlier build evidence; those older builds are not counted as new executions. Physical timing, wiring, sensors, all-board compatibility and original IDE behavioral equivalence remain unverified.

7. **Installer qualification.** NSIS 3.7.3 is preserved byte-for-byte. Knowledge/model hashes match the extracted application. A fresh-data probe of that same extracted binary passes real model/SQLite/Tree-sitter/review restart/Undo and only its own temporary credential roundtrip in the Windows user session. It does not contain this increment's features. Static NSIS inspection finds optional AppData deletion only when the user checks Delete app data and not in update mode, plus non-recursive installation-directory cleanup. No clean VM was supplied/discovered through available host tooling; GUI installation, upgrade, rollback and uninstallation remain **NOT_EXECUTED**. Keep the deletion checkbox unchecked to preserve settings/recovery data. No existing user installation/projects/settings/credentials were changed.

8. **Security findings.** Lockfiles are unchanged; prior full npm audit reports zero advisories for that lock. RustSec primary pages were rechecked for selected Windows maintenance notices (`fxhash`, `paste`, no patched versions) and remaining Linux quick-xml risk. The previous whole-lock report retains 10 advisory records across 8 packages, including aliases/maintenance/platform-only paths; it is not all clean. No new full OSV batch or penetration test was run. Semantic execution reuses containment and restricts drivers/response files; a user-selected executable remains trusted code. Prompt-injection transport tests are mitigations, not proof against actual adversarial models. See 12-dependency-security.md.

9. **Remaining blockers.** A ready authorized model/provider; independent human query authors/reviewers; a disposable clean Windows VM and lifecycle fixtures; actual desktop IPC qualification of semantic navigation; header/unopened-project reference accuracy and broader board/source coverage; physical board measurements; real power/storage/hostile OS-race fault injection; unresolved platform dependency migration. The current optional semantic path was tested on Windows/clangd 23.1.0 only and starts a process per query rather than offering full editor LSP functionality.

10. **Eight original requirements.** All remain **PARTIAL**, with implementation evidence rather than percentages:

| Requirement | New progress | Completion evidence still missing |
|---|---|---|
| System architecture | Existing architecture retained; contained interactive processes and interruption regressions | Cross-platform/system-failure qualification |
| Hybrid RAG | Frozen-set stratified reporting, model retained, blind review gate | Independent judged set and acceptable multilingual accuracy |
| Knowledge-base data | Uncertainties documented; exact offline Block contracts exposed | Complete corpus/metadata/licensing coverage |
| Code-aware chunking | Optional exact-CDB semantic navigation with unsaved buffers and syntax fallback | Desktop IPC/header/all-project accuracy and broader semantics |
| Compiler validation/auto-fix | Phase/revision approval gates retested; tool data cannot authorize application | Real-provider compile/edit/repair quality and wider failures |
| Technology integration | Existing Windows SDK/process/provider/editor paths reused, no new dependency | macOS/Unix execution and platform dependency risks |
| Development phases | Nine-ID typed expansion, two actual SDK builds, source-regression evidence | Generic Blocks/all boards and deferred product work |
| Evaluation/quality | Expanded tests, actual semantic/SDK probes and preserved installer checks | Real LLM, independent annotation, clean VM and physical verification |