AlgoSu ADR

Goal

Wave C: Inject difficulty (difficulty/level) context into AI analysis prompts to calibrate rubrics by tier — Bronze vs Platinum should be evaluated against different expectations.
Wave D: Extend the Programmers crawler to store I/O examples and constraints in dedicated structured fields, then inject them as a <structured_content> block in AI prompts.
Overcome the limitation from Sprint 248 where only problem title and description were injected, giving AI access to the full problem specification.

Sprint 248 completed group analysis problem injection (Wave B) and the Programmers crawler (Wave A), but two gaps remained:
1. No difficulty context: AI evaluated Bronze and Platinum code with identical rubrics — no calibration possible.
2. Full description text injection: Crawled HTML injected verbatim — I/O examples and constraints buried in natural-language description, increasing AI parsing overhead.
An assumption existed that Submission entity already had difficulty/level fields — Wave C pre-check disproved this (migration required).

Rather than querying the Problem entity on each analysis request, copy difficulty/level at submission creation time into the Submission record.
Rationale: avoids an extra Problem Service call during analysis; preserves the difficulty as-of-submission as an immutable snapshot.
Migration: difficulty VARCHAR(20) NULL, level INT NULL columns added to Submission.
ai-analysis _build_difficulty_context(): generates rubric calibration text based on level and difficulty tier (Bronze/Silver/Gold/Platinum/Diamond + Programmers Level 1–5).

Added four structured fields to Problem entity: inputDescription TEXT NULL, outputDescription TEXT NULL, constraints TEXT NULL, examples TEXT NULL.
Extended crawler.service.ts: parsing div.markdown now separates I/O example tables (.example-io) and constraints (<h5>) into dedicated fields — stored independently in the Problem record.
ai-analysis <structured_content> block: injects <input_spec>, <output_spec>, <constraints>, <examples> sub-blocks only when structured fields are present — reuses ADR-030 S-5 prompt injection defense pattern.

_format_examples() injected I/O table cell values directly without _sanitize_problem_field() → risk of breaking prompt boundaries if cells contained <problem_context> tags.
Fix: applied _sanitize_problem_field() to all headers and cell values.
Added 2 tests: normal sanitize + malicious tag blocking.

File	Change
`submission/src/submission/submission.entity.ts`	`difficulty`/`level` nullable columns added
`submission/src/database/migrations/…AddDifficultyLevel.ts`	TypeORM migration
`submission/src/submission/submission.service.ts`	`create()` copies difficulty/level from ProblemServiceClient response
`ai-analysis/src/prompt.py`	`_build_difficulty_context()` added + `build_user_prompt()` parameters extended
`ai-analysis/src/worker.py`	difficulty/level extracted from message fields → passed to `analyze_code()`
`ai-analysis/tests/`	Tests updated

PR #464 squash merge → 6882dc7 (2026-07-15)
Submission: 388 passed · Branches 98.69% ≥ 98%
AI Analysis: 359 passed · TOTAL 99%+
Security patch bundled: multer CVE-2026-5079 — submission service 2.1.1 → 2.2.0 (Sprint 248 patched only problem/ai-analysis; submission was missed)

File	Change
`problem/src/problem/problem.entity.ts`	`inputDescription`/`outputDescription`/`constraints`/`examples` 4 fields added
`problem/src/database/migrations/…AddStructuredContentToProblems.ts`	TypeORM migration
`problem/src/crawler/crawler.service.ts`	Structured parsing (I/O example tables + constraints separation)
`ai-analysis/src/prompt.py`	`_format_examples()` + `<structured_content>` block (I/O + constraints)
`ai-analysis/src/worker.py`	structured content fields extracted + passed to `analyze_code()`
`ai-analysis/src/claude_client.py`	structured data parameter added
`ai-analysis/src/main.py`	structured fields forwarded via Problem Service query
Tests (multiple)	ai-analysis +138 lines, problem +163 lines new cases

Wave D commits: b5b1fbb (Wave D core) · ac70dbd (ruff auto-format) · 87ee8bb (Critic P2 sanitize fix)
AI Analysis: 379 passed · TOTAL 98.89% ≥ 98%
Problem: 213 passed · Branches 96.79% ≥ 96% · Functions 98.61% ≥ 98%
ESLint: Errors 0

Severity	Finding	Fix
P2 example cell sanitize missing	`_format_examples()` — cell values injected without `_sanitize_problem_field()` → malicious tags could break prompt boundaries	sanitize applied to all headers and cells + 2 tests added (`87ee8bb`)