fix(cli): unbreak openkb remove — registry pruning + scoped lint pass (closes #58)

[SeungwookHan]

Summary

Two regressions from #51 surface together on a single openkb remove run but have independent root causes — reported and fixed together because they share one repro, each landed as its own atomic commit.

Closes #58.

Bug 1 — orphan hash entry on legacy documents

After openkb remove reports success, hashes.json still contains the removed doc's entry. Re-running openkb add <same-file> is then wrongly treated as a duplicate via SHA dedup.

Root cause. Commit c504e26 fixed add_single_file to persist doc_name into the registry, but rows ingested before that commit were never backfilled — they still carry only {name, type}. HashRegistry.remove_by_doc_name matched on meta.get("doc_name") == doc_name, which evaluated to None == "<slug>" and silently returned False. The CLI never checked the return value, so the failure was invisible at the surface.

Fix (commit 1). Prune by the already-resolved file_hash (returned by _resolve_doc_identifier) via a new HashRegistry.remove_by_hash(file_hash) -> bool. The hash key is always present, so pruning no longer depends on the metadata shape — and it drops the slug round-trip the CLI never needed.

Bug 2 — fix_broken_links reformatting unrelated wiki pages

Removing a single doc produced a sprawling diff. In the repro, removing one ollama.md produced a 39-file / 1254-line diff; 27 of those were concept pages that never listed ollama as a source.

Root cause. fix_broken_links was built for openkb lint --fix — an explicit, user-invoked wiki-wide cleanup — and strips every dangling [[wikilink]] in the KB. #51 reused it for remove's post-pass, whose declared intent (per the #51 PR description itself) was only to strip links pointing at the just-removed summary and deleted concept pages. The post-remove cleanup intent was scoped; the helper it called was not. That mismatch swept the rest of the KB and silently demoted pre-existing ghost links the user meant to keep (e.g. a hand-written link to a concept they plan to add later).

To be precise: the wiki-wide behavior is correct for openkb lint --fix. This is a #51-side scoping fix, not a change to #49's behavior.

Fix (commit 2, the preferred path from the issue). Add a keyword-only restrict_to parameter to fix_broken_links. remove passes only the union of the concept pages it modified (concept_result["modified"]) and index.md, so the post-remove pass cleans only the dangling links this remove just created. The valid-target whitelist is still computed wiki-wide so legitimate [[concepts/sibling]] links in the scoped pages still resolve. openkb lint --fix is untouched — it omits the parameter and keeps the wiki-wide behavior.

Test contract change

test_cli_remove_lint_cleans_dangling_links previously planted a stray link in concepts/llm.md (an unrelated page) and asserted it was stripped — that test enshrined the over-zealous behavior and couldn't tell scoped cleanup from global. Renamed to ..._in_modified_page and re-targeted at concepts/attention.md (a page this removal actually modifies); the new test_cli_remove_preserves_ghosts_in_unrelated_pages locks in the opposite contract.

Test plan

• tests/test_remove.py — 46/46 pass, incl. test_hash_registry_remove_by_hash, test_cli_remove_prunes_legacy_registry_entry_without_doc_name (E2E Bug 1), test_cli_remove_preserves_ghosts_in_unrelated_pages (E2E Bug 2), test_cli_remove_lint_cleans_dangling_links_in_modified_page.
• tests/test_lint.py::TestFixBrokenLinksRestrictTo — 5 new cases: default behavior, scoped, empty list = no-op, paths outside wiki/ skipped, wiki-wide target set still consulted.
• Full suite: 380 passed, no regressions, no behavior change for openkb lint --fix.
• TDD verification. With only the code patches reverted (tests applied), the 7 new regression assertions fail for the right reasons — AttributeError on remove_by_hash, TypeError on the restrict_to kwarg, the legacy h_legacy entry surviving, and [[concepts/agent-loops]] getting demoted to plain text — and the 2 sentinel tests pass on both sides of the patch.
• Each commit verified in isolation (Bug 1 commit: 96/96 pass with Bug 2 changes stashed).

Commits

1. fix(cli): prune the hash registry by file_hash so openkb remove drops legacy entries — Bug 1
2. fix(cli): scope openkb remove's post-pass lint to the files it touched — Bug 2