fix: stop pointing the agent at an empty episodic log under a memory backend

[AmirF194]

Summary

The Workspace section of the system prompt always advertises {workspace}/user_memory/episodic/episodes.md as the episodic log to grep, in both render.py's identity_text (the live per-turn prompt) and builder.py's hand-duplicated copy (token-size estimation only). That file is written only by the host's legacy MemoryConsolidator. When a MemoryBackend plugin is wired (e.g. EverOS), episodes live in the plugin's own store instead and are surfaced through per-turn recall into the # Memory section (MemorySegmentBuilder), so the workspace file stays empty and the guidance sends the agent to grep nothing, exactly as the issue's own evidence shows (0-byte episodes.md next to 8 populated EverOS episode files).

identity_text now takes a has_memory_backend flag. IdentitySegmentBuilder and build_context_engine already receive backend, so wiring it through is one added parameter on each. With a backend, the Episodic log line is replaced with guidance pointing at the automatic # Memory recall instead of a file path that stays empty; without one (the default), the rendered text is unchanged. ContextBuilder._get_identity now delegates to the shared render.identity_text instead of keeping its own copy of the same text, which is the "unify" half of the issue: the two had already drifted once (builder.py was missing the language directive render.py has).

Type

• Fix

Verification

Commands run (this container, python:3.12-slim + uv sync --frozen --extra dev --dev):

• uv run --extra dev pytest tests/test_segments.py tests/test_context_invariants.py tests/test_default_context_engine.py tests/test_security_untrusted_context.py -v -> 42 passed.
• New regression tests (test_identity_with_backend_omits_stale_episodic_path, test_identity_with_backend_matches_legacy_builder in tests/test_segments.py) fail on main with TypeError: IdentitySegmentBuilder.__init__() takes 2 positional arguments but 3 were given (confirmed by reverting only the source files and re-running against the new tests) and pass on this branch.
• Broader regression sweep: every test file that constructs ContextBuilder/AgentLoop/build_context_engine (29 files, git grep list), minus tests/integration/ and the messaging-channel adapter tests that fail to import for missing optional deps unrelated to this change (dingtalk_stream, lark_oapi, nh3, botpy, slackify_markdown, python-telegram-bot, wecom_aibot_sdk, none installed by --extra dev) -> 402 passed, 22 failed. All 22 failures are in tests/test_runtime_checkpoint_bug2*.py and are pre-existing environment gaps (FileNotFoundError: [Errno 2] No such file or directory: 'git', the python:3.12-slim image has no git binary), unrelated to this diff.
• uv run --extra dev ruff check / ruff format --check on every changed file: clean.
• uv run --extra dev pre-commit run --files <changed files>: all hooks pass (ruff, ruff-format, trailing-whitespace, end-of-file-fixer, large-files, merge-conflict, private-key).

Not verified: an end-to-end run against a live EverOS instance (the issue's own repro environment). The fix is scoped to the prompt-text guidance the issue describes and is covered by the unit tests above; I did not stand up EverOS to confirm the rendered prompt reaches a running agent unchanged elsewhere.

Risk

• Security impact considered: none, prompt text only.
• Backward compatibility considered: has_memory_backend defaults to False and the new backend parameters default to None, so every existing call site (all of which construct without a backend) renders byte-identical text to before.
• Rollback path is clear for risky changes: revert the single commit.

Related Issues

Fixes #122