fix: add behavioral depth criteria to task acceptance and parity checks

[jafreck]

Problem

The task graph builder generated only structural acceptance criteria for migration tasks:

• "All symbols correctly migrated"
• "Call-site signatures match"
• "Target code compiles without type errors"

Meanwhile, the parity verifier enforced behavioral equivalence — checking for hollow implementations, dead dispatch, pass-through codecs, and semantic effectiveness. This disconnect caused the code-migrator to produce implementations that satisfied the stated acceptance criteria (compiled, had the right symbols) but were algorithmically hollow — leading to expensive parity retry loops (7+ attempts per task in the zstd migration).

Root Cause

buildAcceptanceCriteria() and buildParityChecks() in task-graph-builder.ts only generated structural checks. The code-migrator had no upfront signal that it would be held to a behavioral standard, so it optimized for the criteria it was given.

Fix

Both functions now generate behavioral criteria for every task containing functions:

Acceptance Criteria (new)

• Every function body fully implemented — no stubs, TODOs, or placeholders
• Behavioral equivalence: same observable outputs using idiomatic target-language patterns (different types, signatures, and error models are expected)
• Implementation depth matches source complexity — no pass-throughs or synthetic wrappers

Parity Checks (new)

• All source code paths and branches reachable in the target
• No hollow implementations: functions must produce non-trivial, input-dependent output
• Internal call chains wired end-to-end — public entry points invoke the same algorithmic stages

These criteria:

• Apply uniformly regardless of task size (a 30-line hash gets the same bar as a 900-line compressor)
• Use behavioral language ("same observable outputs") not structural language, so idiomatic rewrites are not penalized
• Bridge the gap so the migrator knows upfront what the verifier will enforce

Test Coverage

Added 4 new tests:

• should include behavioral acceptance criteria for tasks with functions
• should include behavioral parity checks for tasks with functions
• should not include behavioral criteria for type-only tasks
• should apply behavioral criteria uniformly regardless of task size

All 48 task-graph-builder tests pass. Full suite: 1498 passed.