chore[bc]: LLM addon interface refactor - remove BaseInference and WeightsProvider

[donriddo]

🎯 What problem does this PR solve?

• LLM addon extends BaseInference and uses the WeightsProvider download layer, so the constructor always needs a Loader even when model files are already on disk.
• Primary GGUF path selection (first shard vs single file) was inlined as a one-liner using a module-level SHARD_REGEX global, with no tests.
• Native C++ event vocabulary is normalized inline in index.js (including the stateful post-finetune TPS skip flag), spreading bridge logic across the class.

📝 How does it solve it?

• Remove BaseInference inheritance; compose createJobHandler() + exclusiveRunQueue() from @qvac/infer-base@^0.4.0 directly.
• Constructor now takes { files: { model: string[], projectionModel?: string }, config, logger?, opts? } with pre-resolved absolute paths.
• Shard streaming via bare-fs.createReadStream directly (no WeightsProvider).
• Extract native event normalization into addon.js (mapAddonEvent), including the skipNextRuntimeStats state used to drop the stale TPS trailer the C++ addon emits after a finetune terminal.
• Extract primary-path selection into a named pickPrimaryGgufPath(files) function exported for testing (with unit tests documenting the .tensors.txt-first ordering and single-file fallback).
• Constructor validation: throws TypeError('files.model must be a non-empty array of absolute paths') for missing/empty input.
• run()-before-load() guard: throws Error('Addon not initialized. Call load() first.') instead of dereferencing null. finetune() already had this guard.
• load() on an already-loaded instance is now a silent no-op (ReadyResource pattern). Call unload() first if you intend to swap weights, then load() again.
• load() is serialized through the same exclusiveRunQueue as run(), finetune(), and unload(), so two concurrent load() calls on the same instance no longer race past the configLoaded guard and leak a native handle.
• Constructor now validates each files.model entry with path.isAbsolute(), plus files.projectionModel when supplied. Relative paths are rejected at construction time.
• Crash-safe shard streaming: partial load failure best-effort-unloads and nulls this.addon.
• unload() clears this.addon and state.configLoaded so post-unload calls hit the guards instead of dereferencing a disposed native handle. pause(), cancel(), and the job-handler cancel closure use optional chaining for the same reason.
• Forwards cacheKey and saveCacheToDisk from runOptions into the text message sent to the addon (from upstream/main's 0.15.0 KV cache API simplification).
• Dependency changes: @qvac/infer-base ^0.3.0 → ^0.4.0, @qvac/dl-base and @qvac/dl-filesystem removed from devDependencies. bare-fs remains a runtime dependency (already present).
• Restored JSDoc on FinetuneValidationSplit.fraction and FinetuneValidationDataset.path that were accidentally dropped during the BaseInference removal merge.
• Fixed docs/architecture.md and docs/data-flows-detailed.md: four occurrences incorrectly stated the "last shard" was the primary path; actual code selects the first shard-regex match.
• Hardcoded shard filenames in the sharded model-loading integration test (was previously regex-derived).
• Removed dead _isSuppressedNoResponseLog filter tied to the old BaseInference._jobToResponse Map; the new architecture cannot emit that message at all, so the wrapped-logger indirection is gone. The user-supplied logger is used directly.

🧪 How was it tested?

• npm run lint (standard) and npm run test:dts clean.
• 31 finetuning unit tests pass after _hasActiveResponse clearing was aligned with the embed pattern (.finally() only; no redundant synchronous clear in _handleAddonOutputEvent).
• New unit tests: test/unit/map-addon-event.test.js (9 tests) and test/unit/pick-primary-gguf-path.test.js (4 tests).
• All 17 integration test files updated to the new constructor shape.
• Examples and benchmarks parse-checked.

💥 Breaking Changes

BEFORE (≤ 0.15.x):

const FilesystemDL = require('@qvac/dl-filesystem')
const loader = new FilesystemDL({ dirPath: '/models' })
const model = new LlmLlamacpp({
  loader,
  modelName: 'Qwen3-1.7B-Q4_0.gguf',
  diskPath: '/models',
  logger: console,
  opts: { stats: true }
}, { ctx_size: '4096', gpu_layers: '99' })

AFTER (0.16.0):

const model = new LlmLlamacpp({
  files: {
    model: ['/models/Qwen3-1.7B-Q4_0.gguf'],
    projectionModel: '/models/mmproj-model-f16.gguf'
  },
  config: { ctx_size: '4096', gpu_layers: '99' },
  logger: console,
  opts: { stats: true }
})

Additional behavior changes:

• In-memory streaming from network sources is no longer supported — the SDK does not use it today (stores models to disk first); can be re-added via a future loader abstraction. Previously possible through the Loader abstraction.
• Second load() is a silent no-op (ReadyResource pattern). Call unload() first if you intend to swap weights.
• downloadWeights, the Loader interface, and loader-based progress callbacks are gone.
• The inherited destroy() is removed; callers use unload() instead.

Version bump: 0.15.0 → 0.16.0. Upstream/main took 0.15.0 for the KV cache API simplification; this refactor lands on the next minor to avoid collision.

🔌 API Changes

// Constructor: single object with files.model (string array) and optional files.projectionModel
// Removed: downloadWeights(), Loader interface, WeightsProvider, BaseInference inheritance, destroy()
// Removed deps: @qvac/dl-filesystem, @qvac/dl-base
// Kept: load(), run(), finetune(), pause(), cancel(), unload()
// Added: getState(), pickPrimaryGgufPath() (exported for unit testing)

[donriddo]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (4/1)
- 1 Team Lead OR Management approval ✅ (2/1)

**Bypass rule:** Triggered (2+ Team Lead approvals (Tier 1 exception)). This PR is approved regardless of tier.

---
*This comment is automatically updated when reviews change.*

[gianni-cor]

/review

[gianni-cor]

/review

[gianni-cor]

/review