Finish the unimplemented surface

[ibuilder]

Nine pages, eleven analytics helpers and one missing column. All of them were 500s that no test covered, because no test had ever requested the page.

Before and after

Walking all 90 GET routes signed in, against a seeded database:

	before	after
200	72	83
500	2	0
unhandled exception	4	0
503	3	0
server errors total	5	0

The missing column

blueprints/project_templates.py:80 set project.template_used on every project created from a template, but no column existed — SQLAlchemy kept it as a transient attribute and discarded it at commit, so the provenance was silently lost on every template-created project. Line 228 then filtered on it and raised AttributeError.

Migration 0003 adds it. upgrade, db check (no drift) and downgrade all verified against a real database.

Nine templates

admin/{edit_user,company_settings,audit_logs,system_status}, azure/{dashboard,configure}, projects/{my_templates,template_preview}, reports/project_report.

Each renders the exact context its view already passed. The details that matter: the audit log iterates logs.items and paginates, because logs is a Pagination and not a list; the role select emits enum names, because the view does UserRole[value]; and the three endpoints I first pointed at admin.* are user_management.* — there are two admin blueprints, /admin and /management, and these pages belong to the second.

All nine verified rendering with real content, 9–46 KB each.

system_status was displaying fiction

It returned a literal dict — database "healthy", "245ms" average response time, 127 requests per minute, 0.2% error rate. Nothing measured, nothing ever changing. An administrator opening it to decide whether the system was in trouble was reading numbers that could not move.

It now measures database round-trip time, cache backend, broker presence, per-integration credential state, and process metrics. Request rate and error rate are omitted rather than invented — nothing records them — with a pointer to /health/metrics.

Eleven analytics helpers

optimize_resource_allocation and generate_project_insights each called a chain of helpers, nine of which did not exist, so both raised AttributeError on their first line of real work.

Implemented against real data. Resource optimisation reports utilisation per resource, names what is over-allocated and by how many units, prices the excess at each resource's own unit cost, and ranks the moves by severity. Portfolio insight reports completion rate, spend against approved budget, throughput across six periods, and DCMA health per project.

Both are deterministic — a test asserts the same data gives the same answer, which is the property a schedule review needs — and neither requires Azure OpenAI.

Where the data cannot support a claim, they say so rather than guessing:

"future_predictions": {
  "available": false,
  "reason": "No projects completed in this window, so there is no rate to project."
}

Benchmarking is against the DCMA 14-point standard the platform already implements, not an invented "industry average".

A correction to my earlier report

The other two helpers were never missing. They were attached to the class at import time with AzureAIPredictiveAnalytics._x = _x, which no AST can see — so the static check reported them falsely, and my claim that they had been "dedented out of the class" was wrong. They are ordinary methods now.

Both did return fixed text — "Project completion rates are stable", "Budget adherence is within acceptable range" — for every company on every request, whatever the data said. Tests assert those exact sentences no longer appear.

tests/test_static_integrity.py now documents that blind spot, so the next finding is treated as a lead to confirm with hasattr rather than a fact.

Guarding it

tests/test_all_routes.py walks the URL map — not a hand-written list — and requests every GET route signed in, failing on any server error. It also fails if a route is skipped for want of a placeholder, so adding a converter cannot quietly drop coverage.

Both KNOWN_MISSING allowlists are now empty.

One test was asserting less than it looked

test_the_downgrade_reverses_cleanly ran a bare db downgrade, which moves back one revision. It only asserted what it meant while 0002 happened to be the head; adding 0003 made it pass through an untested state. It names its target revision now.

Verification

327 passed, 2 skipped locally (was 315). ruff format --check and ruff check clean.

🤖 Generated with Claude Code