api: /maxflow?trace=true ships the whole trace statelessly (closes #36)

[bgent19]

Implements #36. Based on classdemos (not main) because it stacks on #35's api.py changes.

GET /maxflow gains trace: bool = False. When true the whole Edmonds–Karp trace ships as one array — every augmentation, complete enough to draw, with no server-side cursor.

What changed

• Each step is a complete snapshot, never a delta, so backward stepping is an index decrement. EdmondsKarp._snapshot states the entire residual graph and every arc's flow each time.
• Step 0 is prepended — empty path, zero bottleneck, initial residuals — so the uniform residuals from k−1, path from k render rule has a before-state for iteration 1 without building residual graphs in the browser.
• The data layer stopped filtering; the CLI started. Terminals are named __source__ / __sink__ and retained; unbounded residuals ship as null and stay in place, so lane_residuals is one term per hop and a min(...) term can be anchored to its lane. cli.cmd_maxflow narrows at print time.
• flows states every arc, zeros included, closing the trap where a saturated lane was indistinguishable from an absent one. Server-computed, not cap − residual in JS.
• MaxFlowResponse / TraceStepOut document the shape at /docs. trace=false keeps all four existing fields and their values, adding an empty trace array.

Edge identity is deliberately not threaded through the arc generator, per the issue.

Verification

• tests/golden/ pins the --trace lecture text character for character. The goldens were generated from the CLI before the split existed, so they are evidence the narrowing is exact rather than approximately right — a reviewer independently checked out the pre-change commit and diffed; both datasets came back empty.
• New contract-test section covers step 0, the full field set, len(lane_residuals) == max(len(path) - 1, 0) (the spec's formula, guarded for step 0's empty path), terminal retention, null-never-Infinity (parsed with parse_constant so a browser-breaking Infinity cannot slip through), flow coverage of saturated lanes, statelessness, and the /docs schema.
• Two pre-existing core tests in test_maxflow_interdiction.py asserted the old filtered shape and were updated to the new contract, keeping their intent — including the load-bearing edge-order assertions that select which cancellation trace students see.
• Full suite: 203 passed.

🤖 Generated with Claude Code