feat: add extractive_summary workload and pipeline support

[Exohayvan]

Objective

Add a new deterministic local workload, extractive_summary, to extend the AetherMesh prototype from basic text preprocessing into a more useful AI-adjacent task while staying fully offline and replay-verifiable.

Why this matters: this change demonstrates end-to-end useful-work accounting for richer local jobs, by keeping output validation, routing, and contribution crediting in the same deterministic pipeline already used by other workload types.

What changed

• Added extractive_summary to local job execution in the runner with deterministic sentence splitting, local token-frequency scoring, tie-breaking, and output shape (summary, sentences, sentence_count, source_sentence_count, character_count).
• Enforced strict extractive_summary payload validation in runner and validator (text required, max_sentences optional integer 1–10, booleans excluded).
• Wired validator recomputation for extractive_summary so completed outputs are accepted only if they exactly match deterministic rebuild results.
• Added contribution scoring rules for valid extractive_summary outputs with capped integer units and defensive malformed-output handling.
• Added extractive_summary to default capabilities and capability-based scheduling paths so assignment only occurs on nodes advertising support.
• Updated README and example batch manifest to document the new job type, behavior, and scoring model.
• Expanded tests for runner, validation, contribution scoring, scheduler routing, simulation flow, CLI snapshots, and manifest/node-capability behavior to cover determinism, bounds, malformed payload rejection, and audit/accounting paths.

Validation

• Ran full test suite in this workspace: PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python -m unittest discover -s tests -v.
• Result: 318 tests passed, 0 failures.
• No secrets, webhook URLs, or external credentials were touched.

Safety notes

• No force push used.
• No hard reset used.
• Local automation paths are not published in this PR body.

[Exohayvan]

Automated review did not approve this PR

Automated Hermes/GPT-5.5 review decision for PR #54: bad.

Review notes

Ad-hoc verification completed again for the focused changed behavior.

Temporary script:
[local temp path redacted]

What it checked:

• extractive_summary runner boundary behavior:
  • Empty sentence-boundary segments are ignored.
  • Trailing text without terminal punctuation is retained.
  • The output has exactly 2 source sentences with the expected sentence text.
• extractive_summary contribution scorer defensive behavior:
  • Malformed outputs added in tests/test_contribution.py all score 0 contribution units.

Command run:
PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python [local temp path redacted]

Result:
ad-hoc verification passed: focused extractive_summary runner and contribution scorer behavior

Cleanup:
The temporary script was removed successfully.

This is ad-hoc focused verification only, not a claim that the full suite is freshly green.

Safety notes

• Local review log path intentionally redacted.
• No force push used by the automation.
• No hard reset used by the automation.

[Exohayvan]

Closing during PR cleanup: after #51 merged, this extractive_summary workload branch conflicts broadly across the shared workload registry/runner/validation/examples/tests. It should be regenerated on top of current main rather than manually forced through.