Fixing Caching Bug... Again (#78)

* feat: add provider/model formatting and robust model display parsing ✨

**✨ Features**
- Add formatProviderName and formatModelName helpers to modelCapabilities to produce human-friendly provider and model labels.
- Add extractBackendNameFromTooltip in modelPicker to robustly parse backend names from tooltips with vendor fallback.

**🛠️ Refactor**
- Update LiteLLMProviderBase to use the new formatting helpers, construct stable display IDs and unified tooltips, and simplify item details for consistent UI labels.

**🧪 Tests**
- Update model display test expectations and add tests for formatProviderName/formatModelName; also adjust integration test type declarations for consistency.

**🧹 Chores**
- Bump package.json version to 1.6.1-dev1.
- Change post-session-validation script to run "npm run format" and "npm run lint" instead of the `:check` variants.

**📚 Docs**
- Update project-planner prompt to save plans under `.plans/pending` and remove obsolete deliverables text from plan-generation instructions.

* refactor: standardize model display labels and metadata ✨

**✨ Features**
- Add ExtendedModelInformation type and formatModelDisplayLabel helper to centralize model UI labeling and vendor display.
- Update LiteLLMProviderBase to expose rawModelName, vendor, backendName, tooltip and detail so consumers render consistent model info.
- Normalize provider handling and token fields by using providerLower for family derivation and derived.maxInputTokens for maxInputTokens.

**🐛 Fixes**
- Update post-session validation message to recommend running "npm run format" and "npm run lint" (remove :check variants) so commands match current scripts.

**🧹 Chores**
- Bump package.json version to 1.6.1-dev5.

**🧪 Tests**
- Adjust unit tests to validate new display label formatting and backendName/description values.

* chore: release 1.6.1 — update changelog, bump version, deps & dependabot 🚀

**✨ Features**
- Add 1.6.1 changelog entry documenting model display labels, provider metadata, and token normalization.

**🐛 Fixes**
- Document post-session guidance to run `npm run format` and `npm run lint` in the changelog.

**🧹 Chores**
- Bump package.json version to 1.6.1 to mark the release.
- Update PostHog dependencies (`posthog-js` and `posthog-node`) to newer patch versions.
- Tighten Dependabot: switch npm and GitHub Actions updates to weekly and reduce open PR limit to 5.

* feat: 🔁 improve streaming parsing and retry handling

**✨ Features**
- Add buffered parsing for `/responses` tool calls so fragmented tool arguments are assembled before emission. This improves support for LiteLLM response streams and Gemini-style native function calls.
- Preserve deterministic part ordering across thinking, text, tool calls, response metadata, data parts, and finish events. This helps VS Code consume streamed parts more reliably.
- Introduce a retry path for context overflow errors that re-trims prompts with a hard budget override and retries once. If overflow persists, surface a `LanguageModelError` so VS Code can compact the conversation.

**🐛 Fixes**
- Prevent stale or partial tool-call buffers from leaking across aborted streams and new requests. This reduces corrupted tool-call emissions after reconnects or resets.
- Resolve model picker selections by stored model ID instead of label matching alone. This fixes cases where displayed labels include vendor prefixes or other formatting.
- Improve overflow detection with broader error matching for known provider codes and message patterns. This makes retry behavior more consistent across backends.

**🧹 Chores**
- Bump VS Code engine and typings requirements to 1.114.0 and mark the package as preview. This aligns the extension with the newer host API surface.
- Add `.devcontainer/devcontainer-lock.json` to ignore rules and register the post-session validation hook for `SubagentStop`. This keeps local dev artifacts and automation behavior tidy.

* fix: 🐛 drop cache_control metadata from transport and token counting

**🐛 Fixes**
- Add a shared `isCacheControlMimeType` helper and use it to silently drop opaque cache-control parts before they reach outbound LLM payloads.
- Prevent `application/vnd.cache-control+json` variants from being decoded as message text in both V1 and V2 conversion paths.
- Exclude cache-control metadata from V2 token counting so prompt-cache markers do not inflate the budget.
- Update regression tests to verify cache-control parts are stripped while legitimate text and JSON data parts still pass through.

**🧪 Tests**
- Add coverage for bare `cache_control`, `application/vnd.cache-control+json`, and mixed-content messages to guard against metadata leakage.

* fix: 🐛 improve streaming error logging and telemetry

**🐛 Fixes**
- Replace `console.error` with structured logger warnings when tool-call argument parsing fails, and emit a safe empty input fallback so the response stream keeps flowing.
- Add bounded SSE frame parse warnings in `decodeStream`, including payload previews and suppressed logging after repeated bad frames.
- Enrich V2 chat request failures with request ID, model, duration, telemetry failure capture, and failure metric reporting.
- Add unit coverage for tool-call parse fallback, request-failure telemetry, and SSE parse warning behavior.

* feat: ✨ overhaul V2 responses conversion and validation

**✨ Features**
- Add a direct V2-to-responses converter that preserves part ordering, tool calls, and structured tool results without first downgrading to transport messages.
- Serialize tool results as tool messages with support for text, JSON, and data payloads, while keeping adjacent trailing text as its own message when needed.
- Drop cache-control data parts and keep image parts as OpenAI `image_url` content items to better match LiteLLM/OpenAI expectations.

**🐛 Fixes**
- Tighten V2 validation to reject empty messages and require tool calls to be followed by matching tool results in subsequent user messages.
- Prevent flat concatenation of mixed tool-result payloads so structured results remain distinguishable in the emitted OpenAI payload.
- Add coverage for tool-result ordering, structured serialization, and the updated validation behavior.

**🧹 Chores**
- Bump package version from `1.6.2-dev5` to `1.6.2-dev6`.

* feat: 🎉 release 1.6.2

**✨ Features**
- Overhaul the V2 responses pipeline to correctly handle `LanguageModelDataPart`, `LanguageModelThinkingPart`, and other VS Code carrier objects. This improves conversion and validation for response processing.

**🐛 Fixes**
- Improve streaming error logging and telemetry to capture more detail on connection drops and parse failures. This makes stream issues easier to diagnose.
- Drop `cache_control` metadata from transport and token counting to prevent carrier objects from corrupting prompts or inflating token usage.

* fix: 🛡️ drop cache-control carrier parts across streaming and conversion

**🐛 Fixes**
- Drop `cache_control` VS Code data carrier parts in the stream interpreter and emitter so opaque prompt-cache metadata is not re-sent into future LLM input.
- Filter cache-control data parts from V1/V2 tool result serialization and tool-result text collection to prevent metadata leakage while preserving legitimate text output.
- Keep non-cache-control data parts flowing through as before, and add trace logs for dropped carrier parts to aid debugging.

**🧪 Tests**
- Update streaming tests to verify cache-control carrier objects are suppressed and other data carrier objects still pass through.
- Add emitter tests to ensure cache-control parts are not reported to VS Code.
- Add utility coverage for V1/V2 tool-result serialization to confirm cache-control payloads are removed while real tool output remains intact.

**🧹 Chores**
- Remove VS Code test extension auto-installation from `.vscode-test.mjs`.
- Bump the package version to `1.6.3-dev1`.

* Rel 1.6.3 prep