The repo-research-analyst agent previously jumped straight into
open-ended exploration without first grounding itself in the project's
actual technology stack. This led to unfocused research and missed
context that downstream planning agents (like ce-plan-beta) needed to
make sharp decisions about external research.

This adds a structured Phase 0 scan that runs before the existing
research phases:

- 0.1 Root-level discovery: single broad glob to detect ecosystems,
  then selective manifest reads (only files that exist)
- 0.1b Monorepo detection: reuses 0.1 data to identify workspace
  structure and scope subsequent scans
- 0.2 Infrastructure & API surface: conditional checks that skip
  entire categories when 0.1 rules them out (e.g., no API/data layer
  globs for a pure CLI tool)
- 0.3 Module structure: scans internal boundaries within the detected
  scope

The scan output is a Technology & Infrastructure section at the top of
every research summary, reporting languages, frameworks, deployment
model, API styles, data stores, and monorepo layout.

ce-plan-beta's SKILL.md is updated to leverage this structured output
for smarter external research routing -- using detected framework
versions for targeted doc lookups and skipping external research when
local patterns are already established.

Tested with 3 eval scenarios (basic scan, scoped monorepo, absent tech
layer), each with skill vs baseline comparisons:

- Basic scan: 44 vs 68 tool calls, 57K vs 86K tokens (2x faster)
- Scoped monorepo: Phase 0 completed in 4 tool calls
- Absent tech: correctly reported absence of gRPC/GraphQL/REST as
  explicit signal rather than omitting it

Token efficiency was the main iteration focus -- the conditional skip
rules prevent wasted globs when 0.1 already shows the project has no
backend framework, database, or server dependencies.

- Read workspace config files (pnpm-workspace.yaml, nx.json, lerna.json)
  when detected in root listing so monorepo scoping can extract actual
  workspace paths
- Require root listing to show no API-related directories/files before
  skipping API surface checks, preventing false negatives for Go/Node
  stdlib servers
- Check for infra files within scoped monorepo subtrees, not just root

…e probe

- Data layer and API surface now have independent skip conditions. A CLI
  or worker with a database (e.g., Prisma/SQLite) but no HTTP surface
  will still get its data layer scanned.
- Go multi-module detection now adds a conditional */go.mod glob when Go
  directories are visible but no root go.mod was found.