feat: implementation plans + finish match-exhaustiveness & cooperative fiber yield

[MelbourneDeveloper]

TLDR

Adds implementation plans for the compiler's seven remaining partial features, and fully implements two of them — match exhaustiveness with catch-alls, and cooperative yield.

Details

Planning docs — docs/plans/ gains an index (README.md) plus seven actionable plans (0001, 0002, 0004, 0005, 0007, 0008, 0009), each ending in a TODO checklist, covering the compiler features that currently bail or no-op.

Match exhaustiveness — crates/osprey-types/src/pattern.rs:

• is_catch_all is tightened so a capitalised bare binding is treated as a (possibly mis-spelled) constructor rather than a catch-all. Previously any such binding silently disabled the whole exhaustiveness check; now a misspelled variant (e.g. Bleu over Red | Green | Blue) is reported instead of absorbing the missing variants.
• New check_redundant_arms reports dead arms: any arm following an irrefutable catch-all, and duplicate constructor/variant arms.
• A new is_irrefutable (wildcard / lower-case variable only) is used for reachability, distinct from is_catch_all — so a refutable typed-binding arm (n: Int =>) does not wrongly mark a following _ unreachable.
• pattern_ctor_name now counts bare built-in Result variant bindings (Success/Error) toward Result exhaustiveness.

Cooperative yield — crates/osprey-codegen/src/fiber.rs + compiler/runtime/fiber_runtime.c:

• gen_yield previously lowered yield e to plain identity, leaving the C fiber_yield as dead code. It now emits a real call to fiber_yield, forwarding the operand (value and type preserved); the fiber_yield(x) function form shares the same lowering.
• fiber_yield replaces its no-op stub with a real cooperative hand-off: sched_yield() in concurrent (thread-backed) mode, and a deliberate no-op under deterministic mode — which both avoids deadlocking on the runtime_mutex held by fiber_await during deterministic execution and keeps the differential harness byte-for-byte reproducible. (True cross-fiber interleaving under deterministic mode needs stackful context switching; noted as out of scope in spec 0011.)

The 0003 and 0006 plan docs are removed now that they are implemented; spec 0011's status note for yield is updated.

How Do The Automated Tests Prove It Works?

• New osprey-types unit tests (pattern.rs): misspelled_uppercase_variant_is_not_a_catch_all (the Bleu typo is reported), arm_after_a_catch_all_is_unreachable, duplicate_variant_arm_is_unreachable, and lowercase_binding_is_a_legal_catch_all (a genuine variable catch-all still suppresses the missing-variant error). All 100 osprey-types tests pass.
• New must-reject cases: match_misspelled_variant.ospo and match_unreachable_after_catch_all.ospo are rejected by the compiler — the differential harness reports FC_REJECT=56 … FC_OK (ratchet of 12 held; the two new cases are not in the escape list).
• yield codegen is exercised by the existing fibers_channels_yield_select_and_done and yield_without_value_and_let_bound_lambda_materialize codegen tests (covering both the with-value and no-operand paths), and end-to-end by compiler/examples/tested/fiber/fiber_showcase.osp (yield 25, fiber_yield(200), yield streamProcessor(...)), which still matches its .expectedoutput.
• Differential harness: PASS=47 FAIL=0 NOEXP=0, FC_OK — no existing example regressed under the tightened exhaustiveness rules or the new yield lowering.
• Coverage thresholds all hold, including the two changed crates: osprey-codegen 95.6% ≥ 95% and osprey-types 99.1% ≥ 98%. Full-workspace cargo fmt --check, cargo clippy --workspace --all-targets -D warnings, and cargo test --workspace are clean.