Chalie Refinement: Mode Gates, Read Skills & Cleanup

ModeGateService now adds a code_eval directive to system prompts when math or coding modes exceed a threshold, verifying responses before replies.

The system now inverts tool promotion; innate skills are always injected, and external tools must route exclusively through find_tools.

The plan-mode fire threshold was lowered from 0.8 to 0.5 to ensure plan-gated productivity skills remain visible for casual prompts.

The reading capability was heavily streamlined: link extraction machinery was removed, and the output format is now strictly unified to show extracted text only.

The read skill now uses browser-fingerprinted fetch, incorporating canonical format and maintaining cookie jar persistence across redirects.

The introspect innate skill was completely removed from the system, simplifying contextual capabilities in converse mode.

A large amount of dead code and unused services were swept away, removing obsolete prompt chains and various backend services.

Time awareness is improved by surfacing local time timestamps to the LLM via a dedicated TimeFormatterService, preventing raw UTC exposure.

• Lowered plan-mode fire threshold from 0.8 to 0.5 for better productivity tool visibility.

• Removed dead link extraction machinery; read returns extracted text only.

• Unified output format for read operations to [read(…)] structure.

• Innate skills are always injected, removing reliance on deprecated tool promotion via ModeGateService.

• TimeFormatterService now surfaces local-tz timestamps to the LLM.

• Removed the introspect innate skill across its handlers and registry entries.