test: end-to-end behavioural tests for every thesis subcommand

55 new tests covering every command's happy path, error paths, help,
flags, and (for agent commands) prompt payloads + thread persistence.

Commands tested
  - top-level:   no-args-prints-help, --help lists all 9 commands, --version
  - setup:       --help, --non-interactive for anthropic + openrouter, bad
                 provider exits 2, --skip-examples respected
  - init:        --help, --force vs preserve AGENTS.md, outline stub,
                 urls.txt instructions, idempotency
  - status:      --help, runs without key, no workspace mutation, reads
                 persisted thread
  - ingest:      --help, default dir, explicit dir, missing dir exits 1,
                 idempotent skip, works without API key
  - curate/style/write/lint:
                 * parametrised --help check across all four
                 * prompts mention the expected keywords (pending, status,
                   style/samples, STYLE.md, thesis/chapters, section id)
                 * write requires section arg
                 * --thread flag persists to data/.thread (parametrised)
                 * agent errors surface + exit nonzero (parametrised)
                 * persisted thread reused when --thread absent
  - chat:        --help, single-turn then /quit, empty input skipped,
                 /new rotates thread id mid-session, --new flag rotates,
                 agent error surfaces + loop continues, /exit alias

Test infrastructure
  - _StubAgent context manager wraps agent.build_agent so chat tests run
    fully offline
  - ws fixture swaps cli.console for a wide console so Rich tables render
    without CliRunner truncation
  - _record_invoke helper captures (prompt, thread_id) tuples for
    agent-dispatch assertions

Coverage impact
  cli.py        74% → 81%
  TOTAL         82% → 86%
  242 passed, 1 skipped (symlink rights)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: NikitasT2003