- Move synthetic ID generation to GeminiChat stream processing.
- Inject IDs directly into chunk objects to synchronize executor (Turn.ts) and history.
- Implement stateful deduplication during streaming to handle SDK repeats/deltas.
- Use stable synth_${promptId}_${index} format for absolute determinism.
- Standardize Turn.ts to trust and fallback to the same stable ID scheme.

- Refactor ContextGraphBuilder to process history in a single, stateless pass.
- Eliminate turn-to-turn persistent state to prevent mapping drift.
- Simplify ContextGraphMapper to utilize atomic reconstruction.
- Ensure structural resilience to history truncation and reordering.

- Introduce HistoryHardener utility to enforce Gemini API invariants.
- Patch role alternation, user-anchoring, and trailing-user requirements.
- Automatically repair orphaned tool calls with sentinel response nodes.
- Integrate hardener as the final step in context rendering.
- Update tests and snapshots to reflect defensive sentinel turns.

- Update graph nodes to store and serialize thoughtSignatures.
- Fix structural misalignment where signatures were dropped during projection.
- Fix bug in GeminiChat where fixed contents with signatures were discarded.
- Ensure tool calls are correctly linked to thoughts for Gemini-3 models.

- Pivot IR from a reconstructive model to a lossless 1:1 mirror of Gemini Parts.
- Update ContextGraphBuilder to perform 1:1 part-to-node projection.
- Update ContextProcessors (Truncation, Distillation, Masking, Snapshotting) to operate directly on pristine Part payloads.
- Simplify reconstruction logic in fromGraph to preserve original Part objects.
- Refactor all context tests and utilities to align with the flattened structure.
- Add safe Part cloning helpers with justified lint suppressions.
- Update system golden snapshots to reflect improved token estimation efficiency.

…ping

- Remove the logic that dropped orphaned function responses, as it was causing 'function name mismatch' errors when the API state became inconsistent.
- Instead, rely on the existing logic to inject missing responses for calls, which is the safer way to restore API invariants.
- This simplification ensures that we don't inadvertently create state mismatches by removing one half of a call/response pair while keeping the other.

- Re-enable the logic in HistoryHardener to drop 'functionResponse' parts if their preceding 'functionCall' was truncated. This is required by the Gemini API to maintain turn pairing invariants.
- Add ID-based deduplication in 'fromGraph' to ensure each node appears at most once in the history, preventing 'function name mismatch' and duplicate ID errors.
- These changes together ensure that the reconstructed history is both valid and consistent with the API's expectations.

…ardening

- Added 'Node Pinning' in ContextManager to prevent 'in-flight' tool calls (calls without responses) from being truncated during context management.
- Refactored HistoryHardener into a multi-pass validator that strictly enforces role alternation and pairing invariants.
- Fixed 'starts with model' and 'mismatched call/response' API errors by ensuring structural integrity is enforced AFTER all tool pairing modifications.
- Updated golden snapshots to reflect improved sentinel messaging.

- Implementation of globally unique synthetic IDs in GeminiChat and Turn.
- IDs now include promptId, turn timestamp, and a persistent counter to prevent collisions.
- Updated ContextGraphBuilder (toGraph.ts) to prioritize API/Synthetic IDs for node identity.
- Ensured that tool execution results preserve the call's unique ID, enabling 1:1 mapping in the graph.
- Verified with system lifecycle tests.

… observation

- Modified HistoryObserver to always process the FULL history from AgentChatHistory. Previously, it was only processing the incremental PUSH payload, which caused the ContextManager to erroneously prune all existing history nodes.
- Updated renderHistory in ContextManager to correctly identify and protect active tool calls during the synchronous rendering phase.
- These changes ensure that the Context Graph remains stable and only prunes nodes when they actually exceed the configured token budget.

…plication and token overcounting

- Implementation of stable, content-based IDs in 'toGraph.ts' using type-safe salted hashing (content + turn/part indices). This ensures node IDs remain consistent even when history objects are re-created (e.g., during thought stripping).
- Used idiomatic TypeScript type guards to safely handle Part subtypes while complying with strict ESLint rules.
- Added ID-based deduplication in 'ContextTokenCalculator' to ensure each unique node is only counted once toward the token budget.
- Refactored 'appendPristineNodes' in 'ContextWorkingBuffer' to skip nodes already present in the buffer by ID.
- These changes resolve the 'cutting too deep' issue where re-synchronization events caused the history to double in the internal buffer, triggering premature and aggressive truncation/summarization.
- Verified with all 113 context and system lifecycle tests.

- Fixed context graph ID collisions by namespacing tool call and response IDs ('call_' and 'resp_'). This ensures responses are not filtered out as duplicates of their corresponding calls.
- Refined token estimation heuristics to exclude massive 'thoughtSignature' metadata blobs (saving 100k+ phantom tokens in complex sessions).
- Updated image token estimates to Gemini 1.5 standards (258 tokens vs 3000).
- Verified with all context unit tests and system lifecycle golden tests.

- Added 'trace-improvements.md' documenting observability goals.
- Implemented 'Budget Audit' tracing to track max/retained tokens and pressure.
- Implemented 'Protection Audit' with structured reasons (e.g., system_prompt, in_flight_tool_call).
- Implemented 'Transformation Lineage' tracing in Orchestrator to track N->M node replacements and size deltas.
- Implemented 'Estimation Calibration' breakdown (Text, Media, Tool, Overhead) to verify token heuristics.
- All 113 context tests and system lifecycle tests passed.

- Updated 'render.ts' to protect all nodes in the most recent turn (logical episode) rather than just the single last node. This prevents partial pruning of multi-part turns which was causing orphaned tool calls.
- Updated 'toGraph.ts' with robust ID extraction that only uses API IDs if they are valid strings, and added content-based hashing fallback to ensure stability across history re-syncs even for tool parts without IDs.
- Verified with system lifecycle tests.

- Implemented index-independent stable IDs in toGraph.ts using content-based turn salts and occurrence counters. This prevents IDs from changing when history is re-indexed.
- Updated renderHistory to return metadata about whether permanent context management was applied.
- Updated GeminiClient to only overwrite chat history if real management (pruning/summarization) occurred, preventing transient hardening sentinels from poisoning the pristine conversation record.
- Ensures atomic protection of all parts in the most recent turn.
- Verified with system lifecycle tests.

…calls

- Fixed timing bug where tool responses were missing during context rendering by adopting requests into history immediately in 'GeminiClient'.
- Implemented 'refineToolResponses' in 'HistoryHardener' to hoist and re-order tool responses, ensuring strict API compliance and better model follow-up.
- Upgraded 'getStableId' to use SHA-256 for robust collision resistance in large graphs.
- Updated 'hardening-improvements.md' to track long-term stability goals.
- Verified with system lifecycle tests (Scenario 1-3).

- Updated hardening roadmap to reflect the decision to preserve mixed turns (tool responses + text) for better hinting support and role alternation stability.
- Confirmed core structural improvements (re-ordering, hoisting, SHA-256) are complete.

…d 'SILENT_SYNC' to break history re-sync loops.\n- Added Synchronous Pressure Barrier in 'renderHistory' to resolve sync/async race conditions.\n- Centralized node protection in 'ContextManager' to globally pin System Prompt and Recent Context.\n- Implemented render caching using graph-hash to reduce redundant overhead.\n- Refined processor thresholds for increased pruning sensitivity.\n- Fixed core test failures and updated system lifecycle golden snapshots.\n- Updated 'bugs.md' to reflect resolved status for all identified anomalies.