Bounded tool I/O and a smaller default tool surface

[huytg2610]

Motivation

High-volume tools could do substantial work before their output hit the truncation layer — a large file, PDF, or slow HTTP response wasted memory, latency, and context budget. Every turn also exposed the full tool catalog, inflating schema/prompt size for ordinary coding tasks. This moves limits and tool-selection to the source of the data.

What changed

Bounded input

• Read streams files incrementally under byte limits, preserving logical line boundaries and CRLF.
• WebFetch/WebSearch enforce a response byte limit, a bounded redirect chain, identity-only encoding, and one cancellable deadline spanning connect + redirects + body streaming.
• PDF extraction and large-file summarization cap source size, extracted text, page count, chunk output, and reduce-stage input.

Tool profiles

• Default standard profile = common coding tools only; research/orchestration/full are explicit opt-ins.
• The profile is applied consistently across provider schemas, dispatch, prompt, web settings, /config validation, and cache keys. Configs without tool_profile default to standard; full restores the complete surface.

Prompt & cache correctness

• System prompt lists the active tool surface instead of a static catalog.
• Planning guidance shown only when EnterPlanMode/ExitPlanMode are active; tmux guidance only when its tool is active.
• Read-only cache entries invalidated before and after file mutations; output truncation bounded even for very long paths.

Correctness fixes

Sev	Area	Fix
High	files.py · SummarizeLargeFile	Summarize the bounded PDF text instead of erroring on the truncation marker (large PDFs are the tool's main use case).
High	fs.py · _read_logical_line	A CRLF ending on the 8 KiB read boundary swallowed a following blank line and shifted line numbers; now probes only on a bare trailing \r.
Med	files.py · summarize reduce	Skip a failed (None) chunk instead of aborting the merge; report chunks actually merged and warn on incomplete coverage (incl. a clipped last chunk).
Med	web.py · DuckDuckGo parser	Skip HTML void tags (<img>, <br>) in both tag handlers so title/snippet depth counters stay balanced.

Validation

• CI on Python 3.10–3.13; smoke tests on 3.10 and 3.13.
• Regression coverage: bounded reads + CRLF boundaries, oversized/compressed responses, redirect + slow-drip deadlines, PDF/summarization limits, profile filtering + cache isolation, prompt assembly, Web API profile config.
• Fixes verified with targeted reproductions plus the bounded-I/O, summarize, tool-registry, and prompt-assembly suites. compileall and git diff --check pass.