Phase 4: continue gradual mypy typing (next slice of #46)

[subindevs]

Follow-up to #46, continuing after #48 (Phases 1-3: setup, implicit-Optional
fixes, StoreProtocol, and the gently.app.tools.* context-typing cleanup).

Current state (after #48)

With [[tool.mypy.overrides]] stripped, mypy . reports 369 errors across
79 of 300 source files (down from 3002 errors / 225 files at the start of
#46). Breakdown by error code:

Code	Count
union-attr	117
assignment	69
var-annotated	42
attr-defined	40
arg-type	31
return-value	24
index	11
operator	10
misc	9
dict-item	8
other (no-redef, name-defined, valid-type, type-var, override, list-item)	8

By subsystem (top-level package, error count / file count):

Package	Errors	Files
scripts	68	7
gently/hardware	58	9
gently/app	51	6
gently/dataset	47	3
gently/harness	39	16
(remaining: diagnostics, gently top-level, gently/eval, tests, launch_gently.py, gently/ml, gently/organisms, examples)	39	19
gently/core	21	5
gently/mesh	15	3
benchmarks/perception	15	3
gently/ui	12	6
gently/analysis	4	2

(Total: 369 / 79, verified — replaces an earlier version of this table whose
rows summed to ~439 and significantly overstated gently/harness,
gently/app, gently/ui, gently/analysis, and gently/mesh.)

Candidate next targets

Per #46's "follow-up issues per subsystem" approach, a few concentrated
clusters look like good single-PR phases (each needs its own root-cause pass
like #48's Phase 3 did for the context dict pattern):

• gently/dataset/* (3 modules, 47 errors) — aggregator.py (31),
  explorer_server.py (11), embryo_dataset.py (5). The densest cluster:
  47 errors in just 3 files.
• gently/hardware/dispim/* (8 modules, 56 errors) — the device-control
  layer: device_layer.py (14), client.py (14), sam_detection.py (10),
  device_factory.py (10), plus 4 smaller devices//plans/ modules
  (3 + 2 + 2 + 1).
• gently/app/agent.py + gently/app/orchestration/timelapse.py
  (28 + 10 = 38 errors) — the core agent/timelapse loop; may share a root
  cause the way gently.app.tools.* did.
• gently/mesh/* (3 modules, 15 errors) — peer_client.py (13) +
  discovery.py (1) + capability_provider.py (1).

Carried over from #48

• gently/harness/detection/queue.py — 2 of its errors match the Phase 3
  context.get(...) pattern, but it also has 3 unrelated errors
  (var-annotated, an "object" attribute access, and an unsupported
  indexed assignment) that need separate investigation.

Tasks

• Pick one cluster above (or another concentration) as the Phase 4 scope
• Root-cause analysis + short plan (à la Gradually introduce mypy type checking (Phases 1-3) #48's Phase 3)
• Implement, re-measure, remove cleared modules from
  [[tool.mypy.overrides]]
• File the next follow-up issue and repeat until the override list is
  empty

🤖 Generated with Claude Code

[subindevs]

Carry-over from #46: make CI run mypy with project deps installed

Orthogonal to the override-list shrink tracked here, but worth folding into this chain so it isn't lost when #46 closes.

CI and pre-commit currently run mypy deps-less (pip install mypy==2.1.0 only, relying on ignore_missing_imports = true), so mypy . reports clean. But uv run mypy . — what a contributor gets after uv sync, per CONTRIBUTING.md — surfaces 78 errors in 11 files, none of which are in [[tool.mypy.overrides]], and two of which are Phase-3-"cleared" tool modules (gently/app/tools/focus_tools.py, gently/app/tools/calibration_tools.py).

The cause: with real third-party packages installed (anthropic, numpy, …), mypy resolves their actual types instead of falling back to Any, surfacing genuine mismatches the deps-less run can't see. So the current override list (and the 369-error baseline above) is calibrated against the weaker invocation.

Task

• Install the project (deps) in CI before mypy . so the checker sees real stubs, then fix or exempt the 11 affected modules — and re-measure the override-list shrink against the deps-installed run from here on.

This makes mypy meaningfully stronger and aligns CI with what contributors actually run locally. (Original discussion: #46 comment.)