feat: deep research demo with live terminal helix visualisation

[jkbennitt]

Summary

• Adds a live terminal demo (examples/05_deep_research_live/) — real-time ASCII helix visualiser showing agents spiral from exploration to synthesis
• Zero new dependencies, runs offline with a phase-aware mock provider
• Supports --provider anthropic/openai/local for real runs
• Includes integration smoke tests and GIF recording instructions

Relates to #1

Test plan

• ruff check and ruff format pass on all new files
• Integration tests pass (pytest tests/integration/test_deep_research_demo.py -v)
• Full test suite passes
• Demo runs end-to-end with --fast on Windows
• Manual run on macOS/Linux terminal to verify ANSI rendering
• Record a GIF using one of the documented methods

🤖 Generated with Claude Code

[CalebisGross]

Nice work -- the mock provider with phase-aware rotating responses is a smart design, makes the demo instantly usable without keys.

A few things:

Should fix:

• PR description could be trimmed down -- the "What's in the box" / "Demo highlights" sections read a bit like a press release. Just bullets on what's in the PR is fine.
• Redundant condition in helix_visualizer.py around line 609:

  if canvas[explore_row][col].strip() == "" or canvas[explore_row][col] == " ":

  " ".strip() == "" is always true, so the second branch is dead code. Same thing on the analysis_row check below it.
• No Closes #N in the PR body -- is there a tracking issue?

Worth discussing:

• sys.path.insert(0, _example_dir) in both run.py and the test file is a bit fragile. For the test file, a conftest.py in tests/integration/ that handles the path setup would be cleaner than each test file managing it. For the example script it's more acceptable.

Minor / take it or leave it:

• _resolve_provider silently falls back to mock on typos (--provider anthrpic -> mock with a print). Could be worth erroring out instead so users don't accidentally run with mock when they meant to use a real provider.
• _USE_COLOUR is evaluated once at import time -- fine for the demo, just means tests always run in no-colour mode. Not a problem, just noting it.

The _RoundContext dataclass, NO_COLOR/FORCE_COLOR support, Windows UTF-8 handling, and the try/finally shutdown pattern are all solid.

[jkbennitt]

@CalebisGross — addressed all your review feedback:

• PR description trimmed — removed the press-release sections, plain bullets now
• Dead code fixed — removed redundant or canvas[...] == " " on both phase divider checks (L203, L205)
• Issue linked — added Relates to #1
• conftest.py — extracted sys.path setup into tests/integration/conftest.py, test file now uses clean imports

All 4 integration tests pass, ruff lint/format clean. Ready for re-review when you get a chance.

[CalebisGross]

Fixes look good -- dead code removed, conftest extraction is clean, PR description is tidy.

One thing: commits ea78c01 and e0a6717 have identical messages ("polish: lint fixes, smoke tests, improved synthesis, GIF instructions"). Probably a rebase artifact -- worth squashing before merge.