Recommend mixed-pack purchases (buy N of one pack and M of another)

[guplem]

Summary

Closes #26

The list ranked each product on its own and marked one "best option". Now it can recommend a mix of products for one ingredient when that covers the need with less waste, e.g. "3 of A and 2 of B".

• New recommendCombination solver returns a CombinationRecommendation (products + pack counts + waste). rankProducts and the per-product "best option" chips are untouched.
• The solver searches the pack-count combinations across an ingredient's same-unit products and minimizes total bought while covering every cooking event with non-expired food. It is bounded (_maxCombinationVectors = 20000) and deterministic; above the bound it falls back to the best single product.
• It uses per-cooking-event simulation (a mixed container pool), so the mix depends on individual events, not only the weekly total.
• Falls back to a single product when that is already best (and prefers a single product over an equal-waste mix).
• The copied list prints the recommended combination; the card shows a "Best value" banner for real mixes.
• New ADR 0018 records the solver design; indexed in AGENTS.md.

Test plan

• flutter analyze passes
• flutter test test/ passes (629 tests)
• An ingredient where 2 small + 1 large beats any single product recommends the mix
• An ingredient one product covers best still recommends a single product
• Copied list shows the combination

🤖 Generated with Claude Code

[guplem]

Reviewed by an automated pass (rigorous solver review + a polish round). No correctness bugs in the solver; the review's polish items were applied on the branch. Ready for a human check.

What is done well (verified)

• The solver is correct. The search-space bound (each product capped at its solo pack count) is provably safe and never excludes a better mix; enumeration is deterministic; a single product reduces exactly to the existing _simulateProduct; and total-waste ranking is immune to the display-only expiry/over-buy split. The per-event expiry simulation matches the single-product one (same shelf-life boundary, soonest-expiry-first).
• Falls back to a single product when best, and prefers a single product over an equal-waste mix. New ADR 0018 documents the design.

Applied from the review on this branch

• Card banner vs copied list could recommend different mixes (the banner uses the full menu timeline; the copy optimizes per shop trip). Added a short note under the banner so the user knows the copied list splits per trip and can differ. Two widget tests cover it.
• Removed the dead CombinationRecommendation.isViable field.
• Softened ADR 0018's optimality wording: the search is exhaustive and total-bought ranking is exact, but per-vector expiry feasibility uses a documented greedy heuristic not proven optimal in every case.
• Added tests for the two fragile branches: the >20000-combination fallback to a single product, and an infeasible small purchase losing to a larger one due to expiry. Full suite green (633 tests).

Verdict: approved pending human review. CI analyze-and-test gate must be green before merge.

Note for the reviewer: shares waste_optimizer.dart with #37 and #41, so expect merge conflicts when landing more than one.

[guplem]

Rebased onto the updated main (which now has #34, #29, #28) and reconciled. CI is green and the PR is mergeable.

Reconciliation (verified)

• Copy path: _appendIngredientLines now nests the mixed-pack combination inside #34's per-trip amount buckets. It computes recommendCombination for each trip's rounded remaining and prints combinationPackLines, so the copy still matches the page per trip AND shows the waste-minimal pack mix. Confirmed by reading the merged function.
• waste_optimizer.dart: #29's under-buy machinery and this PR's recommendCombination solver coexist (different functions); ProductRecommendation keeps #29's underBuy/shortfall.
• Card: the "Best value" banner (#26), the freeze note (#28), and the under-buy chip (#29) all render together.
• No test expectations were changed. Full suite passes together: 658 tests.

Verdict: approved pending human review. CI analyze-and-test is green and the PR is mergeable now. Reminder: this and #37 both add a multi-product presentation to the card/copy; if you merge both, reconcile the two into one display.

[guplem]

Re-merged after #37 landed on main. CI is green, the PR is mergeable, and the two overlapping features are now genuinely composed (not just merged).

Copy composition (the crux, verified) in buildIngredientCopyLines:

• Runs recommendCombination to get the waste-minimal pack-size mix (#26), sums its PackSelection packs per productEquivalenceKey into packsByKey, then applies distributeEquivalentPacks to spread each key's packs one-of-each across identical variants (#27). Different pack sizes are different keys, so they stay as the combination chose; only identical variants get the one-of-each spread. Both shared helpers are reused, not duplicated. The per-trip amount still drives it, so the copy still matches the page (#34). Nothing of #37's copy behavior was dropped.

Card: renders the freeze note (#28), the "Best value" combination banner (#26), and #37's equivalent-group rows (which host #29's under-buy chip) together.

• Known redundancy for a human to fold later: the banner shows the whole-menu waste-minimal mix, while the per-product rows show per-trip solo counts, so for a real mix the two can show different numbers for the same item. The banner already carries an in-card note that the copied list splits per shop trip, so it is not misleading, but unifying the two displays would be cleaner.

Test change (1, justified): the copy test "non-equivalent products each keep their full solo count" encoded pre-#26 behavior (listing every product at its solo count). #26 deliberately replaces that with a single waste-minimal mix, so it was reworked to "the mix lists each distinct pack size with its own count" - preserving the real intent (different sizes are not merged). Full suite passes together: 698 tests. ADR 0018 updated to describe the composition.

Verdict: approved pending human review. CI green and mergeable. This is the last of the shopping PRs; merging it plus the current #37/#40 completes issues #23-#29 and #34.