fix(merge-chunks): validate untrusted subagent chunk JSON before merging

[TPAteeq]

Problem

graphify merge-chunks concatenates agent-written .graphify_chunk_*.json files into one graph with only a JSON-decode guard around each file. Those chunks are untrusted output from semantic subagents, so an oversized or runaway payload, or a crafted node/edge id containing a path traversal (../../etc/passwd), flowed straight into the merged graph with nothing stopping it. The project already has a security-capping validator for exactly this threat model (#825 — memory exhaustion and chunk-directory escape via crafted IDs), but it was wired only into the skill merge path, not into the in-process merge-chunks command that feeds build_from_json.

Fix

Route each chunk through load_validated_semantic_fragment, which stats the file size before reading it (so a multi-gigabyte chunk can't blow up memory at read time), parses the JSON, and validates the byte/count caps plus the node/edge/hyperedge id charset that blocks path separators and ... An invalid chunk is skipped with a warning (filter semantics — one bad chunk never aborts the merge). This validation is deliberately not wired into build_from_json / load_graph_json: those must keep loading valid, pre-existing graphs whose AST node ids predate the stricter semantic-id charset, and gating them would reject legitimate graphs.

While wiring this in, review caught that the shared validator was too strict in two ways that would have silently dropped whole legitimate chunks, so both were relaxed (a relaxation only — never a new rejection):

• file_type is no longer gated at all, because build_from_json already coerces any value via _FILE_TYPE_SYNONYMS (unknown → concept, extract emits nodes with file_type values outside the validator's allowed enum, producing warnings on every subsequent update #840) and common LLM outputs like markdown/tool/framework would otherwise fail validation wholesale.
• The id charset now allows Unicode word characters, because build's normalize_id deliberately preserves CJK/Cyrillic/accented-Latin identifiers (the explicit path-separator/.. check still blocks directory escape).

The stale module docstring, which named the wrong skills, was corrected.

Tests

tests/test_merge_chunks_validation.py and an update to tests/test_semantic_cleanup.py: path-escape id rejected; malformed shape rejected; valid chunks merge; a file_type: "markdown" / "tool" chunk now merges (regression guard for the data-loss blocker); a Unicode-id chunk merges; and the validator relaxation plus the still-active path-escape guard at the unit level. Full test suite green (3290 passed).

[safishamsi]

Landed on v8 (ships in 0.9.18), cherry-picked so your authorship is preserved. Thanks @TPAteeq — merge-chunks now validates each untrusted chunk (reusing the #825 fragment validator) and skips a bad one with a warning instead of aborting the whole merge or passing a path-escape id through. Small follow-up added on top: non-numeric token counts are coerced so they cannot TypeError the merge after other chunks have merged.