refactor(cli): split cli.py (1705 LOC, 29 commands) into a package — manifesto soft-cap violation

[hadamrd]

Why

src/forge_loop/cli.py is 1705 LOC, 29 commands. The quality manifesto soft-cap for Python modules is 500 LOC. We're 3.4× over.

Direct customer pain:

• Workers dispatched on a CLI ticket have to skim 1700 lines to find the relevant subcommand. Slows every iteration on the CLI surface.
• New contributors trying to read the codebase end-to-end hit cli.py as the second-largest blob (after the runner) and bounce.
• The 29 commands span unrelated concerns (run / status / events / config / brainstorm / manifesto / dashboard / mcp / replay / repos / roles / brief / record-session / pipeline / retry / pause / resume / stop / init / doctor / sweep / ...) — there's no single mental model.

This is the canonical modernization-gated case: split so that future work on axis-1-to-5 features moves faster.

What

Split cli.py into cli/ (package) with one file per concern:

src/forge_loop/cli/
├── __init__.py       # the Typer app — wires sub-typers, keeps `main` entrypoint
├── run.py            # run / pause / resume / stop / retry / record-session
├── observe.py        # status / events / doctor / config
├── brainstorm.py     # brainstorm / manifesto subcommands
├── repos.py          # repos list/disable/enable (multirepo)
├── roles.py          # roles list
├── replay.py         # replay / brief
├── pipeline.py       # pipeline show
├── dashboard.py      # dashboard subcommands
├── mcp.py            # mcp serve
└── init.py           # init / sweep

Acceptance

• forge-loop --help and every forge-loop <cmd> --help produce byte-identical output before/after. CliRunner snapshot tests per command lock this.
• No subcommand drops; the 29 currently-registered commands are still all there.
• Imports flow: cli/__init__.py builds the root Typer app; each module exposes a register(app) function.
• Each new module ≤ 300 LOC.
• Existing tests/test_cli*.py continue to pass without rewrite.

Axis citation

modernization-gated. Unblocks faster iteration on:

• axis:scm-depth (new CLI subcommands for SCM admin operations land cleanly in scm.py instead of bloating cli.py further)
• axis:rbac-multitenancy (RBAC admin CLI lands in a new admin.py)
• axis:ux-quality (operator tour subcommands land in tour.py)

Test plan

• CliRunner snapshot per command (--help output, exit codes, common flag combos). Diff against pre-split snapshots = empty.
• Regression: every existing tests/test_cli*.py runs green untouched.
• Quality-manifesto pin: tests/test_cli_module_size_caps.py asserts every cli/ module ≤ 350 LOC.
• Adversarial: a deliberate typo in a sub-typer registration ⇒ CI fails the snapshot diff (catches regressions).

File pointers

• src/forge_loop/cli.py (delete — replaced by package)
• src/forge_loop/cli/__init__.py (new — facade)
• src/forge_loop/cli/{run,observe,brainstorm,repos,roles,replay,pipeline,dashboard,mcp,init}.py (new)
• tests/test_cli_split_snapshots.py (new — the snapshot harness)
• tests/test_cli_module_size_caps.py (new — manifesto rule pin)