emrg: task prompts: support per-task extra_prompt from tasks.yml config (rant 2026-08-24T21:59:15)

[argszero]

Summary

Support a per-task supplementary prompt (extra_prompt) configured in tasks.yml under a task's config, injected into every round's task prompt so the task understands project-specific contribution rules.

Rant (verbatim): open_source_prompt.md 支持 tasks.yml config 配置补充提示词（per-task extra prompt）— open-source 任务在 tasks.yml 的 config 里可配置补充提示词（如 extra_prompt），随每轮 prompt 注入，让任务理解项目特定的贡献方式。

Changes

• All 5 builtin task templates (evolution_prompt.md, paper_prompt.md, open_source_prompt.md, promote_prompt.md, journal_prompt.md) gain a conditional section right after the "Current State" block:

{% if task.extra_prompt %}
## Task-specific Instructions (extra_prompt from tasks.yml)
{{ task.extra_prompt }}
{% endif %}

• The task Jinja2 context (the task's config dict) already flows through _build_evolution_prompt(); jinja2.Undefined renders the section empty for tasks without extra_prompt — no backward-compat changes needed.
• Test: test_task_extra_prompt_conditional_all_templates renders every builtin template in BOTH states — with extra_prompt configured the text is injected, without it the section is absent (positive + negative validation, emrg: macOS 签名 import 后加私钥存在性检查 — p12 仅证书链时明确报错（R88 根因定位） #455 lesson).
• Agent.md: pytest count synced 1050 → 1051 (new test).

Example

# ~/.emrg/tasks.yml
- name: deepseek-harness-opensource-tasktype: open-sourceconfig:
project: deepseek-harnessextra_prompt: "创建插件（plugin）也是对 deepseek-harness 的贡献，不应被排除在贡献范围之外"

Verification

• pytest tests/ → 1050 passed, 1 skipped
• python -c "from emrg.client.app import run_client" → OK
• python -m emrg --help → OK

[argszero]

✅ LGTM — cycle 221140 (1/3)

Reviewed head d80d59d against master (d68efce): 7 files (+66/−1). Adds a conditional {% if task.extra_prompt %} section to all 5 builtin task templates (evolution/paper/open-source/promote/journal) after their Current State block — the task config dict already flows through _build_evolution_prompt() with jinja2.Undefined, so tasks without extra_prompt render unchanged. Regression test validates BOTH positive (extra_prompt injected) and negative (section absent) states across all 5 templates (#455 lesson). Agent.md pytest count synced. CI green (test + test-windows, run 32737744643).

[argszero]

✅ LGTM — cycle 222434 (2/3)

Reviewed head d80d59d against master (d68efce): 7 files (+66/−1). All 5 builtin task templates gain the identical conditional {% if task.extra_prompt %} section after Current State — uniform, minimal, zero residual output for unconfigured tasks (jinja2.Undefined). Regression test covers both positive and negative states per template (#455). Agent.md count synced to 1051 (doc-count guard #511). CI green (test + test-windows, run 32737744643). LGTM 2/3.

[argszero]

✅ LGTM — cycle 223111 (3/3)

Reviewed head d80d59d against master (d68efce): 7 files (+66/−1). Conditional {% if task.extra_prompt %} section added uniformly to all 5 builtin task templates; dual-state regression test (positive + negative, #455); Agent.md pytest count synced (doc-count guard #511). CI green (test + test-windows, run 32737744643). 3 consecutive LGTMs from cycles 221140, 222434, 223111 — merging.