Own empty-text-block stripping at the wire boundary (third copy of this logic now exists downstream)

[Anarchid]

Follow-up from agent-framework PR #58 review (2026-07-13).

The Anthropic API rejects empty text blocks with 400 "text content blocks must be non-empty", and every consumer that assembles a request from stored history can produce them (tool-only turns, delivery-failure placeholders, silent/skip turns). There are now three downstream copies of the same defense:

1. context-manager's stripEmptyTextBlocks on the compression path (85f8da1, 0.5.9),
2. agent-framework's live-path sanitize in buildActivationRequest (PR #58),
3. and AF's runInferenceWithInjections still needs one (runInferenceWithInjections misses the empty-text sanitize (buildActivationRequest docstring is no longer true) anima-research/agent-framework#62).

Membrane already owns the wire-boundary repair pass for exactly this bug class — orphan tool_use repair (normalizeToolPairs), same-role envelope merging (mergeConsecutiveRoles), the adapter-level block sanitize in buildRequest. Empty-text stripping belongs in that pass: strip empty/whitespace/non-string text blocks, drop messages left with zero content, cover every consumer and every path at once instead of one path per incident per repo.

Field context: one empty block in a compiled window mutes an agent completely — [inference-failed] on every activation until the message ages out (2026-07-10, resident agent).

Downstream copies can then become no-ops to be deleted at leisure; they are all idempotent with a membrane-level strip.

🤖 Generated with Claude Code