Code Farm economy rebalance: market saturation, infrastructure/defense/cosmetics, mastery track, secondary leaderboards, admin eras, underdog bonus and weekly contracts

Every purchase/upgrade path (new and pre-existing) is now race-safe against concurrent requests via atomic conditional SQL updates.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

CLAUDE.md

@@ -283,6 +283,17 @@ Required patterns: every `page.goto(...)`/`page.wait_for_url(...)` MUST pass `wa
 
 **Never run tests unless the user explicitly asks for it.** Not the full suite, not a single file. Validate with a clean import (`python -c "from devplacepy.main import app"`) and per-language manual checks instead.
 
+## Rigorous correctness verification (money, state machines, concurrency)
+
+The persisted test suite (`tests/unit`/`api`/`e2e`) is example-based: it proves the specific inputs someone wrote down behave correctly. It is structurally blind to three classes of bug - a formula wrong at an input nobody tried, an invariant violated only after a long sequence of actions, and a race that only two nearly-simultaneous requests trigger. **Any feature that touches a spendable resource (coins, credits, quotas), a bounded state machine (levels, tiers, counters with a floor/ceiling), or a read-then-write mutation reachable from more than one request path applies this procedure in addition to, not instead of, the normal persisted tests.** A simple CRUD toggle doesn't need it; an economy, inventory, ledger, or scoring feature does. This was built out fully for the Code Farm economy rebalance (see `devplacepy/services/game/CLAUDE.md`, "Every purchase/upgrade is atomic..." - the worked example, including every bug it actually caught) and is the standing procedure for anything shaped like it going forward.
+
+Four escalating layers, run in order, as disposable Python scripts that call the real functions directly against a temp DB (`DEVPLACE_DATABASE_URL`/`DEVPLACE_DATA_DIR` pointed at a scratch file) - not pytest files, unless the user separately asks for persisted tests too:
+
+1. **Property/invariant checks on every pure function, across the full input domain.** For each formula (a reward, a cost curve, a score, a discount), assert the mathematical property it must have - monotonicity, bounds, non-negativity, idempotency, additivity over a partition of its domain - by iterating the real range (hundreds to thousands of values), not a handful of hand-picked spot checks. A monotonicity bug at input 347 is invisible if only 0, 1, and 10 are tried.
+2. **Stateful fuzzing.** Build N independent entities, fire a long randomized sequence of every mutating action across all of them against a live temp DB, and re-check invariants after every single action: balances never negative, counters that should only increase never decrease, levels/tiers never exceed their declared bounds. Catch the domain's expected exception type and continue - the goal is catching invariant violations and *unexpected* exceptions, not asserting every action succeeds. This proves safety (nothing bad happens) but not liveness (legitimate actions actually succeed) - a bug that wrongly blocks a valid action just raises a well-formed expected exception and sails through a pure fuzz test undetected. Layer 3 exists because of exactly this gap.
+3. **Concurrency: prove races are closed, never infer it from luck.** Any read-then-write mutation reachable from more than one worker process (`uvicorn --workers N`) is a TOCTOU race until proven otherwise. Test with **real separate OS processes**, not threads in one process - `dataset` gives each thread its own pooled connection, and enough threads exhausts that pool and produces `database is locked` noise that is a test-harness artifact, not a finding about the application. Set up genuinely fresh, production-representative state before racing - **never pre-seed or zero a column that the real code path leaves unset/NULL.** This was the single most expensive mistake made building this procedure: a first race-safety pass "passed" only because its own setup script had artificially pre-zeroed columns that a real fresh row leaves as SQL `NULL` - silently hiding the exact bug the fix was supposed to prevent (`NULL = 0` evaluates to `NULL`, not true, in a SQL `WHERE` clause; any column not written in the row's original `INSERT` needs `COALESCE(column, 0)` in every later precondition and every arithmetic `SET`, not a bare comparison). Fire many concurrent attempts at the same resource, assert the exact right number succeed, and verify the final state matches the hand-computed expected total exactly (currency spent, levels advanced) - not just "the others were blocked." Fix a real race with a single atomic conditional SQL statement at the exact chokepoint (`UPDATE ... SET ... WHERE <precondition>`, checked via the driver's real `rowcount` - `dataset`'s wrapped `db.query()` does not expose it, use `db.executable.execute(sqlalchemy.text(...), params).rowcount` inside `with db:`), not a client-side lock or an optimistic-locking library this codebase doesn't otherwise use.
+4. **Static analysis beyond `py_compile`.** A clean `python -m py_compile` and a clean `from devplacepy.main import app` only prove syntax and module-load order - neither catches a missing import inside a function body, which is a runtime `NameError` invisible until that exact line executes. Run `pyflakes`/`ruff check` on every touched file before calling a change done; it is nearly free and catches an entire class of bug that layers 1-3 can each individually miss if they don't happen to exercise the broken line.
+
 ## Feature workflow
 
 Every feature in DevPlace is **one data source fanning out into several consumers**. DevPlace has no isolated change - internalize this before editing. A single handler in `routers/{area}.py` is simultaneously the **four faces of one route**:
@@ -300,7 +311,7 @@ A new public read almost always needs all four. The cardinal failure mode is cha
 4. **View layer.** Extend `base.html`; page CSS in `{% block extra_head %}`, page JS in `{% block extra_js %}`. Reuse partials and template globals. JS is ES6 modules, one class per file, instantiated on `app`.
 5. **Agent + docs layer (the most-forgotten step - do not skip).** `services/devii/actions/catalog.py` - add an `Action` tool if a user could ask Devii to do this (`requires_auth=False` for public reads). `docs_api.py` - every public/auth endpoint gets an `endpoint()` entry with params and a `sample_response`. `routers/docs/pages.py` `DOCS_PAGES` for prose pages. SEO (`seo.py`) - `base_seo_context` and any JSON-LD for a new public page; add to `routers/seo.py` sitemap if indexable.
 6. **Document.** `README.md` (product-facing, routes/config/dependencies/user-visible features), the relevant nested `CLAUDE.md` (any new mechanic, helper, pitfall, or cross-layer wiring), this root `CLAUDE.md` only when a NEW architectural rule or workflow step is introduced (not per feature).
-7. **Validate.** Confirm a clean import (`python -c "from devplacepy.main import app"`), check each touched language manually, grep for em-dash in touched files (neither the character nor its HTML entity - use a hyphen). **Never run the test suite unless the user explicitly asks.** Write new tests in the matching tier/path when asked, following the required patterns above.
+7. **Validate.** Confirm a clean import (`python -c "from devplacepy.main import app"`), check each touched language manually, grep for em-dash in touched files (neither the character nor its HTML entity - use a hyphen), and run `pyflakes`/`ruff check` on every touched file. **If the feature touches a spendable resource, a bounded state machine, or a read-then-write mutation reachable from more than one request path, also run the four-layer procedure in "Rigorous correctness verification" above** - it is the default for that shape of feature, not an optional extra. **Never run the persisted test suite unless the user explicitly asks.** Write new tests in the matching tier/path when asked, following the required patterns above.
 
 Failures at any implementation step block the workflow - never skip a failed step.