Close remaining doc staleness from PR #189 work

Three small gaps caught in a final audit of the §-doc and
the issues doc:

1. feature-investigation §1 'What java-llama.cpp already covers':
   the Session row still read 'single-threaded'. PR #189 added
   per-Session locking; relabel to 'thread-safe per instance'
   so the §1 table matches the §2.6 status block.

2. feature-investigation §4 'Suggested rollout order': all ten
   entries listed there are SHIPPED now. The section is preserved
   as a historical artifact (useful for comparing original effort
   estimates against what actually shipped, especially §2.1 which
   came in much smaller than the L estimate and §2.10 which had
   to be reverted), but a header note now states up front that it
   is no longer a roadmap.

3. docs/history/49be664_open_issues.md #113 LoadProgressCallback:
   the entry said FIXED in PR #188 and stopped there, but the
   PR #188 implementation forgot the forward declaration in
   jllama.h, which made the JNI symbol C++-mangled and produced
   an UnsatisfiedLinkError on the first call to the feature.
   PR #189 commit 36d8862 added the missing declaration and the
   feature is now actually functional (LoadProgressCallbackTest
   passes on the current CI). Add a 'Subtle issue resolved by
   PR #189' paragraph so a future reader does not think 'PR #188
   shipped this, why did it fail on the first run after merge'.

No code changes; documentation only.

Author: claude

docs/feature-investigation-llama-stack-client-kotlin.md

@@ -36,7 +36,7 @@ T-shirt sizes:
 | Reactive Streams `Publisher<LlamaOutput>` token stream   | ✅ (§2.3) |
 | `completeBatch` / `chatBatch` parallel dispatch          | ✅ (§2.4) |
 | Typed `Usage` / `Timings` / `CompletionResult`           | ✅ (§2.5) |
-| `Session` helper (single-threaded)                       | ✅ (§2.6) |
+| `Session` helper (thread-safe per instance)              | ✅ (§2.6) |
 | `AutoCloseable` iterator + cancel polish                 | ✅ (§2.7) |
 | Per-request `setJsonSchema` + `completeAsJson<T>`        | ✅ (§2.8) |
 | Typed `TokenLogprob` in `LlamaOutput`                    | ✅ (§2.9) |
@@ -411,6 +411,17 @@ web frameworks.
 
 ## 4. Suggested rollout order
 
+> **Historical note (kept for context).** All ten items below are now
+> SHIPPED &#x2014; the original §2.1 / §2.2 / §2.5 / §2.7 / §2.3 / §2.4 /
+> §2.8 / §2.9 / §2.6 / §2.10 sequence was delivered across PR #188 and
+> PR #189. §2.10 ships **cooperative only**; the M-effort sub-token
+> follow-up was attempted twice and reverted (see the postmortem in the
+> §2.10 entry above for why). §2.1 ended up much smaller than the
+> original L estimate once the upstream OAI chat path was found to
+> already accept `image_url` blocks. This section is preserved so the
+> original effort estimates can be compared against what actually
+> shipped; it is no longer a roadmap.
+
 1. **§2.1 Multimodal (L)** — biggest capability gap, isolated subsystem.
 2. **§2.2 Typed Chat model + tool calling (M)** — foundational; other
    features (usage, logprobs, async) all return / accept these types.

docs/history/49be664_open_issues.md

@@ -190,8 +190,10 @@ vs. total, whether the file is the weights file, whether it is a download or
 disk load) via a `Consumer<LLamaLoadProgress>` callback passed to the
 `LlamaModel` constructor. Intended for showing a progress bar to end users.
 
-**Status in fork:** FIXED in PR #188 (commit `70df324`). New
-`LoadProgressCallback` functional interface (single method
+**Status in fork:** FIXED in PR #188 (commit `70df324`), with a
+**follow-up symbol-export fix in PR #189 (commit `36d8862`) that the
+original feature needed to actually be callable** on Linux/macOS.
+New `LoadProgressCallback` functional interface (single method
 `boolean onProgress(float progress)`; return `false` to abort).
 New constructor overload
 `LlamaModel(ModelParameters, LoadProgressCallback)` plumbs the
@@ -207,6 +209,22 @@ bytes, weights vs download flag) is NOT exposed — only the float —
 because `llama_model_params.progress_callback` itself only emits the
 float; richer fields would require an upstream API change.
 
+**Subtle issue resolved by PR #189 `36d8862`.** The PR #188 implementation
+forgot to add the `loadModelWithProgress` forward declaration to
+`src/main/cpp/jllama.h`. The `jllama.cpp` translation unit `#include`s only
+that header (not the freshly-generated `net_ladenthin_llama_LlamaModel.h`
+that `javac -h` writes during `mvn compile`), so the C++ compiler treated
+the function definition as a regular C++ function and emitted a name-mangled
+symbol (`_Z57Java_..._loadModelWithProgress...`). The JVM looked up the
+plain unmangled name and threw
+`UnsatisfiedLinkError: 'void net.ladenthin.llama.LlamaModel.loadModelWithProgress(...)'`
+the first time `LoadProgressCallbackTest#receivesProgressUpdates` exercised
+the code path. Adding the seven-line forward declaration in `jllama.h`
+restored `extern "C"` linkage and the test now passes
+(`LoadProgressCallbackTest: Tests run: 3, Failures: 0, Errors: 0` on the
+current CI run). The neighbouring plain `loadModel` symbol was unaffected
+because its declaration was already in `jllama.h`.
+
 ---
 
 ## #112 — Qwen 3 model does not load