Fix Kotlin script and C++ retrieval; release 0.3.2

README.md

@@ -44,7 +44,8 @@ No model provider is required by core.
 
 Local stdio MCP servers already provide repository navigation and separate signed
 context/graph tools. Claude Code and Codex are first-class target clients; actual
-Codex use is recorded, while Claude model/tool acceptance remains open. See the
+Codex use and a first Claude read-only task are recorded; full Claude lifecycle
+and routine coding acceptance remain open. See the
 [client and language plan](docs/CLIENTS-AND-LANGUAGES.md) for current capabilities,
 deeper Rust/TS/Kotlin work and the provider-authentication boundary.
 

RELEASE_EVIDENCE.md

@@ -4,6 +4,27 @@ Latest implementation and adoption results are in the
 [dogfood execution ledger](docs/DOGFOOD-EXECUTION.md). The dated snapshots below
 remain historical evidence; internal pilot acceptance is separate from G0–G4.
 
+## 0.3.2 shipment checks, 22 September 2026
+
+Navigation and exact source packets now accept `.kts`, `.cc`, `.cxx`, `.hh`,
+`.hpp` and `.hxx`, matching the broad scanner's existing lexical coverage.
+Syntax-aware planning remains TS/JS-only and Dart remains unsupported.
+See the [language support matrix](docs/LANGUAGE-SUPPORT.md).
+
+Local Node 24.21.0 checks passed 402 tests, independent package installation,
+both unchanged benchmark gates, navigation stdio smoke, 22 scale tests and the
+10k postings probe. Added coverage checks each suffix's search, exact source
+lines/hashes, stale rejection and refresh, policy exclusion, unsupported planning
+and retrieval through installed stdio MCP. Commit-specific CI, public artifact
+checksums and download verification accompany the GitHub release.
+
+npm authentication still returns HTTP 401; registry publication is outstanding.
+Upgrade using matching GitHub tarballs or the source build, then reconnect the
+MCP server: refreshing its source index does not reload its implementation.
+These checks do not establish new Claude lifecycle acceptance or measured
+whole-task/subscription/cash savings. The first actual Claude audit and its
+incomplete host-cost coverage are recorded in the execution ledger.
+
 ## 0.3.1 shipment, 22 September 2026
 
 This patch adds an installed, read-only `doctor` command that verifies the exact

docs/CLIENTS-AND-LANGUAGES.md

@@ -47,15 +47,18 @@ contracts, keeping unsigned navigation and signed evidence clearly distinguished
 Actual Codex calls have succeeded, including this review. The recorded Claude
 Code health check connected, but a previous Haiku attempt returned HTTP 429 for
 a weekly limit before tool use. This is historical evidence, not a current quota
-check. Claude task acceptance remains unverified. See
+check. On 22 September, Claude Code 2.1.278 used all four tools from an installed
+0.3.1 package for a [language coverage audit](LANGUAGE-SUPPORT.md). The read-only
+task was accepted after host corrections; actual-client edit/stale/refresh,
+restart/disable and consequential task qualification remain open. See
 [the pilot record](DOGFOOD.md) and [navigation contract](LOCAL-NAVIGATION.md).
 
 ## OS1 client and retrieval assignments
 
 | Task | Owner / where | Model and effort | Acceptance |
 | --- | --- | --- | --- |
-| **OS1a — Claude Code alongside Codex. Next** | Context integration owner, Node tools, public setup docs and isolated client fixtures | Flash/thinking off for bounded docs/fixtures; Terra/medium integration; **Claude Sonnet 5/medium** as the first real Claude executor; **Claude Opus 5/high** for one consequential source-backed task after basic acceptance | A pinned package launches in both clients against an explicit root; real tool discovery, status/refresh/search, bounded evidence, edit/stale/refresh, restart and disable work. Record actual client/model/effort and accepted task, not just Connected |
-| **OS1b — Sufficient evidence through MCP. Implemented locally; live client qualification remains** | Context implementation owner, shared packet module plus `packages/context-tools` adapters/tests | Terra/medium implementation; Sol/high boundary review; Flash/off for scoped fixtures | `repository_packet` returns bounded exact ranges or complete selected TS/JS blocks with provenance and stale rejection. Installed-tarball stdio retrieval, edit/refresh and restart pass. Reconnect existing clients to discover the new tool; actual Claude use remains open. No shell helper required by the consuming model; no arbitrary path escape |
+| **OS1a — Claude Code alongside Codex. Partial: read-only audit accepted** | Context integration owner, Node tools, public setup docs and isolated client fixtures | Flash/thinking off for bounded docs/fixtures; Terra/medium integration; **Claude Sonnet 5/medium** as the first real Claude executor; **Claude Opus 5/high** for one consequential source-backed task after basic acceptance | A pinned package launches in both clients against an explicit root; real tool discovery, status/refresh/search, bounded evidence, edit/stale/refresh, restart and disable work. Record actual client/model/effort and accepted task, not just Connected |
+| **OS1b — Sufficient evidence through MCP. Shipped; first Claude use recorded** | Context implementation owner, shared packet module plus `packages/context-tools` adapters/tests | Terra/medium implementation; Sol/high boundary review; Flash/off for scoped fixtures | `repository_packet` returns bounded exact ranges or complete selected TS/JS blocks with provenance and stale rejection. Installed-tarball stdio retrieval, edit/refresh and restart pass. Reconnect existing clients to discover the new tool; first Claude read-only use is recorded, while lifecycle qualification remains open. No shell helper required by the consuming model; no arbitrary path escape |
 | **OS1c — Graph questions through MCP** | Context graph owner, same tools adapter; follows OS3/OS4 evidence integration | Terra/medium; Sol/high review of scope/provenance | Bounded repository symbol/neighbour/path queries return precise source and derivation metadata in both clients; do not mistake the existing signed-cache graph tools for this capability |
 | **OS1d — Other client surfaces** | Context integration owner, portable examples and clean-client acceptance | Flash/off for docs, Terra/medium review; user's chosen model for the actual client | Claude Desktop is separately qualified after Claude Code; other MCP clients follow demand. Never infer their acceptance from Claude Code or Codex |
 
@@ -80,6 +83,10 @@ client support does not require repeating every benchmark on every model.
 
 ## Current language depth and priorities
 
+The [selected suffix audit](LANGUAGE-SUPPORT.md) records exact navigation, broad
+scan, build-packet and plan-packet outcomes for Kotlin, C/C++, Dart and TypeScript,
+with source references and explicit test gaps.
+
 These findings come from this repository's source and tests, not a new scan of
 every ForgeSworn project. The captured Heartwood pilot confirms Rust, TS/JS and
 Kotlin inputs. Python also implements this project's new snapshot helpers.

docs/DAILY-USAGE.md

@@ -213,7 +213,7 @@ same-model task comparisons. Pricing estimates and invoice attribution are
 different evidence. Account identifiers group usage; no subscription charge is
 invented or multiplied by the number of tasks.
 
-Remaining FS2 work: real Claude export acceptance, client-version coverage,
+Remaining FS2 work: wider client-version coverage,
 complete everyday host/worker/review captures and convenient weekly report
 rendering. Explicit task boundaries and worker-helper receipt imports are available. The current tool handles explicit exports immediately;
 automatic session discovery, provider authentication and telemetry are excluded.
@@ -230,6 +230,29 @@ metadata was unavailable and all whole-task coverage assertions were false.
 This is a partial Context **development** window, not an accepted Heartwood task,
 a baseline, a complete cost of this implementation or evidence of savings. No
 real Claude export was present in the selected Context export directory; the
-Claude adapter has synthetic fixture acceptance only. The active Heartwood
+Claude adapter had synthetic fixture acceptance only at that point. The active Heartwood
 session was not inspected or modified. Private evidence is recorded in the
 [execution ledger](DOGFOOD-EXECUTION.md).
+
+## First real Claude export acceptance
+
+On 22 September 2026, Claude Code 2.1.278 completed a read-only Context language
+coverage audit using an independently installed 0.3.1 MCP package. Its frozen
+transcript contained 19 assistant entries representing seven unique message IDs.
+The existing importer removed 12 identical usage copies, found no conflicts and
+reconciled exactly with the CLI result: 166,442 total input tokens (130,617 cache
+reads, 35,811 cache creation and 14 fresh input) plus 5,322 output tokens. Cache
+creation's nested breakdown was not added again. No adapter code change was needed.
+
+The observed model was `claude-sonnet-5`; medium effort was requested in the launch
+command, but transcript effort remains unknown to the adapter. The CLI separately
+reported 1,310 thinking tokens within output; this adapter deliberately leaves
+reasoning unknown for its supported transcript shape. The result's list-price
+cost is not an invoice or subscription charge.
+
+The audited document was accepted after host source-review corrections. Only the
+Claude execution counters are complete here: Codex preparation, integration and
+review are not fully captured, and review duration is unknown. Task coverage stays
+incomplete. This validates one real Claude export shape and read-only task use;
+it does not establish whole-task cost, normal-session lifecycle acceptance or
+savings. See [the audit](LANGUAGE-SUPPORT.md) and [ledger](DOGFOOD-EXECUTION.md).

docs/DOGFOOD-EXECUTION.md

@@ -926,3 +926,79 @@ The local release-check blocker is resolved. Doctor and benchmark changes remain
 uncommitted and unpublished; this is not new CI, registry or client acceptance.
 Next adoption gate: a real Claude source-packet task and usage import. The existing
 Heartwood owner retains its application task; no consumer checkout was changed.
+
+## 22 September 2026 — First Claude task and real export import
+
+After the 0.3.1 shipment, an independent installation of its downloaded GitHub
+tarballs passed `doctor` against the Context checkout at `bb83fa3`. Claude Code
+2.1.278 then ran a useful read-only language coverage audit using that installed
+MCP server. The actual model was `claude-sonnet-5`, with medium effort requested.
+The existing Max subscription login was used; no new API credential was supplied.
+One client invocation completed successfully, with no reported provider refusal,
+tool error or permission denial and no host replay. Heartwood was not inspected or modified.
+
+The invocation supplied an explicit MCP binding, `--strict-mcp-config`, the four
+allowed repository tools, no built-in tools, and disabled skills for this bounded
+audit. It preserved saved client settings. This establishes real Claude Code
+print-mode use with an explicit binding, not acceptance of every interactive
+configuration. [CLI flag reference](https://code.claude.com/docs/en/cli-reference).
+
+The model called status once, refresh once, search eight times and packet once.
+The source-backed result needed host corrections: it falsely called `.kts`
+absent from broad scanning, misstated the number of packet suffixes, described
+the planner rejection incorrectly and omitted `.hh` from its proposed alignment.
+The reviewed [support matrix](LANGUAGE-SUPPORT.md) is accepted after those
+corrections; the original draft remains partial, retained privately. No second
+model pass was needed. The audit identifies a concrete next task: align Kotlin
+script and C++ suffix eligibility across navigation and exact source packets,
+with per-suffix tests and no broader parsing claim.
+
+The completed transcript was copied from this explicitly created session only.
+The existing daily importer accepted its real shape without code changes: 19
+assistant entries became seven unique request records, with 12 identical copies
+removed and no conflicts. Imported totals exactly match the CLI result:
+166,442 input tokens (130,617 cache reads, 35,811 cache creation, 14 fresh input),
+5,322 output tokens, 171,764 combined. The separate CLI thinking count is included
+in output; imported reasoning and effective effort remain unknown. Reported
+list-price cost is not subscription billing and is not treated as cash spend.
+
+Private launch, transcript, original answer, review, reconciled receipt and task
+summary are retained outside Git. Only the observed Claude execution is fully
+accounted here. Codex preparation, repair/integration and review usage are not
+fully captured; review duration and whole-task cost remain unknown. Coverage
+assertions explicitly retain those gaps. This is Context development, not an
+accepted Heartwood application task, a controlled comparison or evidence of
+token/subscription/cash savings.
+
+OS1a and FS2 now have first real Claude read-only tool-use and export evidence.
+Actual-client edit → stale → refresh, restart/disable, a consequential task and
+complete everyday host/worker/review receipts remain open. Eight searches before
+one packet are an observation to review on further tasks, not a demonstrated
+regression or causal savings claim. Documentation links and source assertions
+were checked; no production code or language eligibility changed in this slice.
+
+## 22 September 2026 — Kotlin script and C++ retrieval alignment (0.3.2)
+
+The language audit's six concrete gaps are fixed: `.kts`, `.cc`, `.cxx`, `.hh`,
+`.hpp` and `.hxx` are eligible for navigation and exact packets. Broad scanning
+already recognised them; syntax planning remains TS/JS-only. No new parser or
+semantic-resolution claim is made. The support matrix now documents this state.
+
+The M4 worker endpoint was occupied by another live session, so no inference was
+dispatched or interrupted. The host implemented and reviewed the small allowlist
+change and regression tests. The first focused run failed seven assertions on
+test error wording; the repair checks actual stale status and the existing
+refresh-required error. All 26 focused tests then passed. Whole host usage and
+review cost remain incomplete; no worker-token or monetary saving is claimed.
+
+Local shipment checks passed 402 tests, independent installed-package smoke,
+both unchanged benchmark gates, navigation stdio smoke, 22 scale tests and the
+10k postings probe. Per-suffix MCP tests cover search, source hashes/lines,
+checkout/commit provenance, stale rejection and refresh, Git-ignore exclusions
+and unsupported plan requests; the installed smoke retrieves all six suffixes.
+Dart remains explicitly unsupported. Publication/CI evidence belongs to the
+0.3.2 release, separately from these local checks; npm auth returned HTTP 401.
+
+Existing servers need a reconnect after the upgrade. No Heartwood checkout,
+configuration or live job was changed. Real Claude lifecycle qualification,
+routine whole-task receipts and measured benefit remain the next adoption gates.

docs/GETTING-STARTED.md

@@ -27,8 +27,8 @@ matching core version:
 mkdir context-install
 cd context-install
 npm init -y
-npm install --ignore-scripts /absolute/path/to/forgesworn-context-0.3.1.tgz \
-  /absolute/path/to/forgesworn-context-tools-0.3.1.tgz
+npm install --ignore-scripts /absolute/path/to/forgesworn-context-0.3.2.tgz \
+  /absolute/path/to/forgesworn-context-tools-0.3.2.tgz
 ```
 
 For that installation, use

docs/LANGUAGE-SUPPORT.md

@@ -0,0 +1,64 @@
+# Source coverage: Kotlin, C/C++, Dart and TypeScript
+
+Updated for 0.3.2 on 22 September 2026. The original audit used 0.3.1
+(`bb83fa3`); the six suffix eligibility gaps found there are fixed in 0.3.2.
+This is a selected capability audit, not a complete language manifest. Source
+allowlists establish eligibility; ignores, selection policy and resource bounds
+can still exclude eligible files.
+
+| Suffix | Repository navigation | Broad source scan | Exact `build` packet | Syntax-aware `plan` packet |
+| --- | --- | --- | --- | --- |
+| `.kt` | Yes | Kotlin lexical extraction | Yes | Unsupported |
+| `.kts` | Yes | Kotlin lexical extraction | Yes | Unsupported |
+| `.cpp` | Yes | C++ lexical extraction | Yes | Unsupported |
+| `.cc`, `.cxx` | Yes | C++ lexical extraction | Yes | Unsupported |
+| `.h` | Yes | Classified as C; lexical extraction | Yes | Unsupported |
+| `.hh`, `.hpp`, `.hxx` | Yes | C++ lexical extraction | Yes | Unsupported |
+| `.dart` | Excluded | Unsupported | Rejected | Unsupported |
+| `.ts` | Yes | Not handled by this scanner; separate TS/JS scanner exists | Yes | TypeScript syntax blocks |
+
+The source contracts are
+[navigation's suffix allowlist](../packages/context-tools/src/repository-navigation.ts),
+[the broad scanner's language map and lexical rules](../packages/context-tools/src/broad-source-scan.ts),
+[the TS/JS scanner](../packages/context-tools/src/source-scan.ts), and
+[packet source/plan allowlists and guards](../packages/context-tools/src/source-packet.mjs).
+
+
+Navigation provides exact-token line lookup. Broad scanning uses regular
+expressions and labels inferred relationships; it is not a language parser.
+The dedicated TS/JS scanner and packet planner use TypeScript syntax trees, but
+do not establish type-checked semantic resolution. Exact build packets copy
+selected lines with provenance; they make no syntax-completeness guarantee.
+Unsupported plan suffixes are explicitly rejected by the planner guard before
+TypeScript parsing; rejection is not a parser failure on those languages.
+
+## Test evidence and remaining gaps
+
+- [MCP packet tests](../packages/context-tools/src/repository-packet-mcp.test.ts)
+  exercise each added suffix, plus uppercase `.HPP`, through exact search, packet
+  lines and hashes, checkout/commit provenance, stale rejection, generation
+  replacement and retrieval after an edit. Git-ignore exclusions still suppress
+  search and reject packets. Syntax planning still rejects these suffixes.
+- [Broad scanner tests](../packages/context-tools/src/broad-source-scan.test.ts)
+  cover declarations and lexical provenance for all six suffixes and uppercase
+  `.HPP`. Dart remains explicitly unsupported in scanning and MCP retrieval.
+- [Installed package smoke](../test/context-package-smoke.mjs) retrieves all six
+  suffixes through the installed stdio MCP server outside the workspace.
+- Existing packet tests retain TS/JS syntax support and path/policy boundaries.
+  These tests establish source retrieval and lexical coverage, not compiler-level
+  understanding or complete evidence for an arbitrary development task.
+
+Dart extraction and deeper Kotlin/C++ parsing remain separate work. A running
+MCP server needs a reconnect after upgrading its implementation; refreshing the
+source index alone does not load new server code. Measure benefits on actual
+accepted tasks; no token or subscription saving follows from suffix coverage.
+
+## Qualification provenance
+
+Claude Code 2.1.278 used the independently installed 0.3.1 repository MCP package
+for this audit, with `claude-sonnet-5` and requested medium effort. All four tools
+were used. Host source review corrected a false `.kts`-absence claim, an incorrect
+suffix count, planner-rejection wording and the incomplete proposed C++ suffix
+set before accepting this document. The original draft is retained privately.
+See [the execution ledger](DOGFOOD-EXECUTION.md) for usage and remaining client
+acceptance limits. This is not a measured savings comparison.