`--update --compare-to` never regenerates invalidated modules in whole-repository documentation mode (always ends in `IncompleteGenerationError`)

[kiboook]

Summary

When a repository is small enough that module clustering is skipped
("whole-repository documentation mode" — see the "Skipping LLM module
clustering ... using whole-module documentation mode" log line),
--update --compare-to <commit> correctly detects which sub-module
doc is affected by a change and deletes it, but never regenerates it.
The run always ends with:

✗ Documentation generation finished but these module docs are missing: <module>.md
IncompleteGenerationError

This is 100% reproducible, not transient — confirmed by tracing the actual
code path (details below), and confirmed twice against two different repos.

Root cause

In codewiki/src/be/documentation_generator.py,
generate_module_documentation() computes the processing order from
first_module_tree.json:

processing_order = self.get_processing_order(first_module_tree)

But first_module_tree.json is only ever written by the LLM clustering
path (cluster_modules(), in the _run_module_clustering stage). For a
repo that goes through whole-repository mode, that file stays {}
forever — even though a prior whole-repo agent run can still insert real
sub-modules into module_tree.json via
generate_sub_module_documentation_tool.

Sequence:

1. First generate run: whole-repo mode, the agent creates e.g.
   module_a.md, module_b.md. module_tree.json now has 2 entries;
   first_module_tree.json stays {}.
2. A file under module_a changes. generate --update --compare-to <sha>
   runs. _invalidate_affected_modules() (in cli/commands/generate.py)
   correctly matches the change to module_a using module_tree.json,
   deletes module_a.md (and overview.md, since any invalidation also
   invalidates the overview).
3. generate_module_documentation() loads module_tree.json
   (non-empty → 2 entries) so it takes the if len(module_tree) > 0:
   branch, but processing_order (from the still-empty
   first_module_tree) is []. The regeneration loop runs zero
   iterations — module_a is never reprocessed.
4. It falls through straight to building the repo overview, which logs
   Module docs not found at .../module_a.md, and then
   validate_generated_docs() raises IncompleteGenerationError.

Minimal reproduction

No secrets/private code needed — a tiny throwaway repo reproduces this
every time:

src/auth/auth.service.ts, src/payment/payment.service.ts,
src/notification/notification.service.ts, src/app.ts — each a small
class with a couple of methods (~20 leaf components total, well under
the default 36,369-token clustering threshold, so whole-repo mode is
guaranteed).

git init && git add -A && git commit -m "baseline"
# <baseline_sha>

codewiki config set --provider openai-compatible --api-key <key> \
  --base-url <url> --main-model <model> --cluster-model <model> \
  --fallback-model <model>
codewiki generate --output ./docs
# → module_a.md / module_b.md / module_c.md / overview.md created,
#   module_tree.json has 3 entries, first_module_tree.json is {}

# add one function to auth.service.ts, commit again
git add -A && git commit -m "add a function to auth service"

codewiki generate --output ./docs --update --compare-to <baseline_sha>
# → fails every time with IncompleteGenerationError

I have the exact 4 source files used and I'm opening a PR with this fix,
the fixture, and a regression test.

Suggested fix (implemented + tested in the linked PR)

Fall back to module_tree for the processing order only when
first_module_tree is empty — this only changes behavior for the
whole-repo-mode case; the normal clustered path is untouched, since
first_module_tree is non-empty there.

processing_order = self.get_processing_order(first_module_tree)
if not processing_order and module_tree:
    processing_order = self.get_processing_order(module_tree)

Verified against the repro above: the affected module regenerates,
unaffected modules are correctly skipped ("✓ Module docs already
exists"), and the overview is rebuilt referencing all modules.

[anhnh2002]

Thanks for raising this. I'll check asap.