When a user cancels a tool call by pressing Esc twice, the assistant
message with tool_calls was remaining in the conversation history but
no tool response followed. This caused 400 Bad Request errors from
model providers on the next message.

The fix uses the parent context (ctx) instead of the request context
(genCtx) when updating messages during tool call tracking and tool
result creation. This ensures that even if the request is cancelled
mid-stream, the message updates and creations succeed.

Additionally, the error message was updated from 'Tool execution
canceled by user' to 'Error: user cancelled assistant tool calling'
to better indicate the cancellation occurred during the assistant
tool calling phase.

Fixes #529
Fixes #1206

Replace invalid "finish" key name with "enter" in the OAuth success
state key binding. "finish" is not a valid key name in the key package,
while "enter" is consistent with other key bindings in the codebase
and matches the help text shown to users.

* Closes #2495

These ENVs can be used by tools to detect if they are being executed
by Crush, via the Bash tool.

Add detailed documentation to improve code review experience:
- Explain scoring weights and their rationale
- Document the ranking strategy and workflow
- Clarify path hint detection heuristics
- Add examples to helper functions
- Document test cases with their intent

Implement custom ranking algorithm for file completions to provide
better matching results based on user input patterns.

Changes:
- Add intelligent path hint detection to distinguish between filename
  and filepath queries
- Implement weighted scoring system that balances fuzzy matching with
  prefix/contains bonuses
- Support both full path and basename matching with appropriate weights
- Add depth penalty adjustment based on query type
- Select match highlights based on weighted contribution
- Add comprehensive test coverage for ranking logic

- Add DialogID field to file picker command in commands.go
- Add FilePickerID case handler in openDialog function
- Remove TODO comment and complete implementation

5MB is way too much. Many models simply don't have the context windows
to handle that. 1MB seems to works well.

Worth noting that:

* For the view tool, the model will be instruced to read in pieces
* For the fetch tool, the content will be truncated with a note

To reproduce (I tested with Kimi K2.5), use the following prompt:

    read this: https://developers.openai.com/api/reference/resources/responses

This release migrates away from `wazero` in favor of `wasm2go`, which
should bring better performance, specially on startup time.

For more information, see:

* https://github.com/ncruces/go-sqlite3/releases/tag/v0.33.0
* https://github.com/ncruces/go-sqlite3/discussions/361

Reported by Grype.

* Reverts 479ca3f1157153af58a826cdff01cb2c535f0cf6
* Ref https://github.com/contributor-assistant/github-action/pull/174